Paper deep dive
Generative Retrieval for Unsupervised Text-Based Person Search
Mang Ye, Yucheng Ji, Yang Bai, Min Cao, Siyuan Chai, Bo Du, Min Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/20/2026, 2:39:31 AM
Summary
This paper introduces GTR+, a two-stage framework for unsupervised Text-Based Person Search (TBPS) that generates pseudo-text descriptions from unlabeled images and uses them for retrieval training. The generation stage employs a tiered framework (base, intermediate, advanced) using Multimodal Large Language Models (MLLMs) to create fine-grained, diverse captions. The retrieval stage utilizes an adaptive confidence-weighted learning framework with a Gaussian Mixture Model to handle noisy pseudo-texts. Additionally, the authors contribute LargeFine-Person, a large-scale dataset for unsupervised TBPS pre-training.
Entities (8)
Relation Signals (7)
LargeFine-Person → iscontributedby → GTR+
confidence 95% · Beyond that, we also contribute LargeFine-Person... constructed by... generating corresponding textual descriptions using our proposed tiered description generation framework.
GTR+ → solves → Text-Based Person Search
confidence 95% · We propose GTR+, a two-stage generation-then-retrieval framework... for unsupervised TBPS
GTR+ → mitigates → noisy pseudo texts
confidence 92% · to mitigate the impact of noisy pseudo texts, we develop an adaptive confidence-weighted retrieval learning framework.
GTR+ → employs → Gaussian Mixture Model
confidence 90% · We model image-text pairs as clean or noisy using a Gaussian Mixture Model
GTR+ → uses → Multimodal Large Language Models
confidence 90% · The base tier leverages an automated question-and-answer mechanism... MLLM to self-query
LargeFine-Person → derivedfrom → LUPerson
confidence 85% · The dataset is constructed by collecting and refining person images from LUPerson
LargeFine-Person → derivedfrom → LPW
confidence 85% · The dataset is constructed by collecting and refining person images from LUPerson [17] and LPW [18]
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Text-based person search (TBPS) aims to retrieve images of a target person from a large image gallery based on a given natural language description. Most existing methods rely on supervised learning with manually annotated image-text pairs. In this paper, we explore unsupervised TBPS, with only unlabeled images. We propose GTR+, a two-stage generation-then-retrieval framework. In the generation stage, we introduce a tiered description generation framework designed to produce fine-grained and stylistically diverse textual descriptions through a three-tier sequential process. The base tier leverages an automated question-and-answer mechanism to generate basic visual attribute descriptions; the intermediate tier enhances fine-grained detail using an inter-sample contrastive mechanism; the advanced tier further enriches textual diversity via a stylized expansion mechanism. In the retrieval stage, to mitigate the impact of noisy pseudo texts, we develop an adaptive confidence-weighted retrieval learning framework. We model image-text pairs as clean or noisy using a Gaussian Mixture Model, calibrated by real-time image-text similarity and static text generation probability from the prior stage, yielding adaptive sample weights during training. Beyond that, we also contribute LargeFine-Person, a large-scale TBPS dataset with high-quality, fine-grained, and diverse textual annotations, enabling a practical and generalizable TBPS pre-training benchmark under unsupervised setting. Experiments on multiple TBPS benchmarks demonstrate the effectiveness and generalization of both GTR+ and LargeFine-Person. Code is available at: this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.12965v1
- Canonical: https://arxiv.org/abs/2609.12965v1
Trouble viewing inline? Open PDF directly →
Full Text
101,379 characters extracted from source content.
Expand or collapse full text
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE1 Generative Retrieval for Unsupervised Text-Based Person Search Mang Ye, Senior Member, IEEE , Yucheng Ji, Yang Bai, Min Cao, Siyuan Chai, Bo Du, Senior Member, IEEE , Min Zhang Abstract—Text-based person search (TBPS) aims to retrieve images of a target person from a large image gallery based on a given natural language description. Most existing methods rely on supervised learning with manually annotated image-text pairs. In this paper, we explore unsupervised TBPS, with only unlabeled images. We propose GTR + , a two-stage generation-then-retrieval framework. In the generation stage, we introduce a tiered description generation framework designed to produce fine-grained and stylistically diverse textual descriptions through a three-tier sequential process. The base tier leverages an automated question-and-answer mechanism to generate basic visual attribute descriptions; the intermediate tier enhances fine-grained detail using an inter-sample contrastive mechanism; the advanced tier further enriches textual diversity via a stylized expansion mechanism. In the retrieval stage, to mitigate the impact of noisy pseudo texts, we develop an adaptive confidence-weighted retrieval learning framework. We model image-text pairs as clean or noisy using a Gaussian Mixture Model, calibrated by real-time image-text similarity and static text generation probability from the prior stage, yielding adaptive sample weights during training. Beyond that, we also contribute LargeFine-Person, a large-scale TBPS dataset with high-quality, fine-grained, and diverse textual annotations, enabling a practical and generalizable TBPS pre-training benchmark under unsupervised setting. Experiments on multiple TBPS benchmarks demonstrate the effectiveness and generalization of both GTR + and LargeFine-Person. Code is available at: https://github.com/Flame-Chasers/GTR. Index Terms—Text-based person search; unsupervised learning, image captioning, person re-identification. ✦ 1 INTRODUCTION Text-based person search (TBPS) [1], [2], [3] involves retrieving images of a specific individual from an image gallery based on textual descriptions. This task shares sim- ilarities with classical person re-identification [4], [5] and image-text retrieval tasks [6], [7], yet presents unique chal- lenges of its own. In contrast to person re-identification that relies on image-based queries, TBPS offers a more accessible solution with free-form text as queries, but faces challenges in cross-modal learning due to differences between visual and textual data. Compared to general image-text retrieval, TBPS focuses on identifying individuals, which requires attention to fine-grained details and presents greater chal- lenges due to increased intra-class variability (e.g., varying appearances of the same individual) and reduced inter-class variability (e.g., visually similar individuals). Despite these challenges, TBPS has gained attention in recent years for its potential in real-world applications, such as surveillance for locating suspects or missing persons. The current mainstream research in TBPS [8], [9], [10] focuses on developing robust modality-invariant represen- tations through sophisticated cross-modal alignment frame- works. While these approaches improve performance, they •Min Cao and Yucheng Ji are with the School of Computer Sci- ence and Technology, Soochow University, SuZhou, China (e-mail: mcao@suda.edu.cn, 20254227067@stu.suda.edu.cn). •Mang Ye, Yang Bai, and Bo Du are with the School of Computer Science, Wuhan University, Wuhan, China (e-mail: yemang@whu.edu.cn). •Siyuan Chai is with Zhipu AI, Beijing 100089, China. •Min Zhang is with Harbin Institute of Technology, Shenzhen, China. •Min Cao is the corresponding author and this work is dominated at Soochow University. •This work was supported by the National Natural Science Foundation of China under Grants No. 62476188 and T2541022. ... Themanwearsa whiteT-shirtand black... ... Thewomeniswearing ablackslewithablack pantsdsad... Amanwithblack shorthairiswearing black Amanwearsagray jacket,blackpants, and...... Themanhasshort hairandorangeand dark... Sheiswearingawhite T-shirt,bluejeansand apairofpinkshoes,... ImageText (a) Image (b) ... Fig. 1: Illustration of (a) canonical TBPS with supervised image-text data, (b) TBPS rely solely on image corpora. heavily depend on supervised image-text data (i.e., person images with manually annotated descriptions) for training, as shown in Fig. 1 (a). This reliance presents challenges for real-world applications, where annotating large-scale person images with textual descriptions is labor-intensive and time-consuming. Therefore, this work focuses on un- supervised TBPS, eliminating the need for annotated text corpora and relying solely on image corpora for effective retrieval, as shown in Fig. 1 (b). For this, we present a basic two-stage solution, generation-then-retrieval (GTR + ), which firstly generates pseudo text descriptions corresponding to the person images for remedying the absent annotation, and then trains a retrieval modal in a supervised manner. Generation. Given the person images, our goal is to generate their paired pseudo textual descriptions that accu- rately depict the fine-grained details with diverse sentence structures. A simple approach is to use pre-trained Vision- Language Models (VLMs) [11] for direct image captioning. However, as shown in Fig. 2 (a), these models often fail to produce satisfactory results due to their pre-training on arXiv:2609.12965v1 [cs.CV] 11 Sep 2026 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE2 VLM Amanwithblackvest andgreenpants. Human-annotated Texts Predefined Templates Finite Attribute Set Amanwithblackshort hairiswearingablack sleevelessvestandgreen pants.Hehasredhead phonesaroundhisneck andisholdingsomething inhishand. Themanwithblackshort hairiswearingblackvest, greenpantsandholding somethinginhishand. MLLM (a) (b) (c) Fine-grained Enhancement Intermediate Tier MLLM Stylized Expansion Advanced Tier MLLM Attribute Extraction Base Tier MLLM Fig. 2: Illustration of the text generation for a given per- son image. (a) Direct image captioning using a pre-trained vision-language model. (b) Description generation via mul- timodal large language model with auxiliary information. (c) The proposed tiered description generation framework. general image-text pairs, limiting their ability to describe the nuanced features specific to person images. Some stud- ies [12], [13], [14], [15] leverage Multimodal Large Language Models (MLLMs) for fine-grained description generation. However, these methods often rely on predefined tem- plates, finite attribute set, or annotated texts as guidance. Consequently, the generated textual descriptions provide limited information, restricting both the richness and gran- ularity of the outputs, as shown in Fig. 2 (b). To address these limitations, we propose a tiered description genera- tion framework which constructs fine-grained and diverse textual descriptions through a self-contained, multi-tier pro- cess (Fig. 2 (c)). Specifically, it consists of three sequential tiers. 1) Base tier. Instead of relying on fixed templates, we design an automated question-and-answer mechanism that enables MLLM to self-query and extract fine-grained visual attributes (e.g., gender, clothing type, clothing color, accessories). This process aims to generate a basic textual description that captures the essential appearance charac- teristics of the person. 2) Intermediate tier. We develop an inter-sample contrastive mechanism to further capture subtle visual details. By comparing the target image with the hard-negative image, MLLM is prompted to highlight both similarities and differences, enhancing the granularity and informativeness of the descriptions. 3) Advanced tier. Building upon the results of the first two tiers, we propose a stylized expansion mechanism to enhance textual style diversity. The mechanism leverages MLLM to automatically enrich descriptions in various styles, guided by the pre- viously generated texts. Overall, our framework produces fine-grained and stylistically diverse textual annotations in a tier-wise manner, which serve as effective supervision for training TBPS model. Retrieval. This stage adopts generated image-text pairs to employ supervised training of the retrieval model. The retrieval model can, in principle, be integrated with any existing TBPS method [3], [16]. However, it is important to note that these TBPS methods are typically designed to train on manually annotated and well-aligned image-text pairs. In contrast, the constructed pairs used as training data in our approach may contain noise due to potential halluci- nations and inaccuracies introduced during the preceding generation stage. To mitigate the impact of noise on the re- trieval model training, we propose an adaptive confidence- weighted retrieval learning framework. Confidence scores, reflecting image-text correspondence reliability, are adap- tively incorporated into the loss to calibrate error propaga- tion, assigning higher weights to more reliable pairs and enhancing their influence on learning. Confidence scores are computed using a Gaussian Mixture Model (GMM) that separates image-text pairs into clean and noisy compo- nents based on their similarity distribution. As training pro- gresses, the similarity is updated adaptively—incorporating both real-time image-text alignment and prior text gener- ation probability—enabling the GMM to iteratively refine sample reliability. This progressive refinement sharpens the clean-noisy distinction, yielding a more robust and accurate retrieval model. In addition to proposing a GTR+ solution for un- supervised TBPS, we also contribute a large-scale TBPS dataset with fine-grained and diverse textual descriptions, named LargeFine-Person. The dataset is constructed by col- lecting and refining person images from LUPerson [17] and LPW [18], followed by generating corresponding textual descriptions using our proposed tiered description gen- eration framework. As a result, we construct a total of 3.6 million image-text pairs. LargeFine-Person enables a practical and generalizable TBPS pre-training benchmark under unsupervised setting. A comparison between the proposed LargeFine-Person dataset and other pre-training TBPS datasets [13], [14], [15], [19], [20] is presented in Table 1. LargeFine-Person presents superior data qual- ity—especially in terms of textual descriptions—attributed to the proposed tiered description generation framework. In summary, our main contributions are as follows: 1)We propose a tiered description generation frame- work, integrating automated question-and-answer generation, inter-sample contrastive generation, and stylized expansion generation to produce semanti- cally rich and diverse captions for person images. Based on this, we construct LargeFine-Person, a large-scale dataset of 3.6 million high-quality, fine- grained image-text pairs. 2)We propose an adaptive confidence-weighted re- trieval learning framework that mitigates the impact of noisy pseudo-texts via self-adaptively weighting image-text pairs based on the confidence scores. These scores are iteratively refined using a Gaussian Mixture Model that integrates image-text similarity and text generation probability. 3)We conduct comprehensive experiments on multi- ple TBPS benchmarks. The results demonstrate the effectiveness and generalizability of our proposed GTR+ solution and LargeFine-Person dataset in un- supervised settings. A preliminary version of this work has been presented IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE3 Image DomainText Domain Datasets Source#ImagesDenoiseSource#Texts#Words/Text#Adj.#NCs.Multi-granularity MALS [20]Generation1.51M ✓Human.1.51M272.727.97✗ LUPerson-T [19]LUPerson [17]1.30M ✗Auto.1.30M274.395.86✗ LUPerson-MLLM [13]LUPerson [17]1.00M ✗Auto.4.00M293.478.62✗ SYNTH-PEDES [14]LUPerson-NL [21] + LPW [18] 4.79M ✓Human.12.14M244.817.08✗ HAM-PEDES [15]SYNTH-PEDES [14]1.00M ✗Human.2.00M265.667.54✗ LargeFine-Person (Ours)LUPerson [17] + LPW [18]1.20M✓Auto.3.60M619.6315.27✓ TABLE 1: Comparison between our pre-training dataset (LargeFine-Person) and existing pre-training datasets for TBPS. ’Denoise’ indicates whether low-quality images are filtered out from the dataset. ’Human.’ specifies that the generation of texts depends on manually annotated inputs, while ’Auto.’ denotes that the text generation is fully automated without manual intervention. ’#Words/Text’ represents the average number of words per text, reflecting the richness of information conveyed. ’#Adj.’ denotes the average number of adjectives per text, reflecting the fine-grainedness of the descriptions. ’#NCs.’ denotes the average number of noun chunks per text, indicating the completeness of the attribute coverage. ’Multi- granularity’ indicates the presence of texts with varying levels of detail, reflecting the diversity of the textual data. Benefiting from the proposed tiered description generation framework, our dataset demonstrates advantages in both diversity and fine-grained detail. at ACM M 2023 [12]. This paper introduces the follow- ing significant advancements: (1) At the task level, while prior work predominantly focuses on TBPS without parallel image-text data—utilizing separately collected image and text corpora where textual data is still required, we focus on the more challenging and practical task of unsupervised TBPS, where only an image corpus is available. (2) At the model level, we propose a generation-then-retrieval framework, which extends and improves upon previous work in both text generation and retrieval modeling. In the text generation stage, we design a novel tiered description generation framework, which comprises three progressive components: an automated question-and-answer mecha- nism to capture basic attributes, an inter-sample contrastive mechanism to encode fine-grained details, and a stylized expansion mechanism to enrich the expressiveness of the generated descriptions. This tiered framework facilitates the generation of descriptions that are both semantically rich and stylistically diverse. It significantly improves upon previous work, which typically relied on manually designed prompts and external text corpora for image-to-attribute extraction and attribute-to-text conversion—often leading to monotonous outputs. In the retrieval learning stage, we introduce an adaptive confidence-weighted learning framework that enhances robustness to noisy data. Unlike previous work that used a static text generation probability as the confidence score, our method dynamically adjusts confidence scores during training, enabling them to adapt to the learning dynamics. This allows for more accurate weighting of training samples and better noise estimation, ultimately improving retrieval performance. (3) At the data level, leveraging the proposed tiered description genera- tion framework, we additionally contribute a pre-training dataset LargeFine-Person, which enables a generalizable pre-training benchmark under unsupervised setting. These resources are expected to advance the application of TBPS in real-world scenarios. 2 RELATED WORK 2.1 Text-based person search Existing TBPS approaches can be categorized into two main paradigms based on their primary focus: cross-modal alignment and representation learning. The first line of research focuses on cross-modal alignment, aiming to map visual and textual features into a shared embedding space. Early methods relied on global feature alignment [22], [23], [24], while later works progressed to multi-granularity cor- respondences [16], [25], [26], [27], [28] and incorporated external tools—such as human parsing, pose estimation, and NLTK [29]—for fine-grained feature alignment [30], [31], [32]. Recent advances have achieved progress in self- adaptive semantic alignment [9], [33], [34], [35]. The second paradigm focuses on learning modality-invariant represen- tations that generalize across modalities [36], [37], [37], [38], [39]. For instance, Wu et al. [40] introduced color-reasoning sub-tasks to enhance fine-grained cross-modal associations, while Wang et al. [41] proposed a color deprivation module to reduce over-reliance on color and learn more robust features. Recent efforts [2], [3], [38], [42], [43], [44], [45] have utilized large vision-language models like CLIP [46] and multimodal LLMs to enhance fine-grained retrieval. Several studies [10], [47] further incorporated multi-round interac- tions and contextual scene information [48] for better query- image alignment. However, these methods still rely heavily on costly, labor-intensive manually annotated image-text pairs, limiting their scalability and real-world applicability. Recently, researchers have begun to recognize the an- notation burden inherent in traditional TBPS and have explored less annotation-intensive paradigms. Several stud- ies [49], [50], [51], [52], [53] focused on weakly supervised TBPS, where image-text pairs are available without explicit identity labels. Gao et al. [54] explored a semi-supervised setting with limited labeled pairs and abundant unlabeled images, while Jing et al. [55] studied cross-domain adap- tation from labeled source domains to unlabeled targets. Though reducing annotation effort, these methods still re- quire partial supervision. In contrast, fully unsupervised TBPS methods [1], [12], [56] required only unannotated images. They typically generated pseudo-text annotations followed by noise-robust learning. However, the perfor- mance is limited by: (1) template-based text generation yields limited and single-perspective descriptions, and (2) existing noise modeling strategies consider only local text- image similarity. In this work, we solve unsupervised TBPS IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE4 that addresses both limitations via three-tiered text genera- tion and adaptive confidence-weighted retrieval learning. 2.2 Unsupervised Vision-Language Tasks Considering that the paired image-text data are extremely labor-intensive to collect, significant efforts have been dedicated to advancing unsupervised learning in vision- language tasks, including unsupervised vision-language pretraining [57], [58], [59], [60], unsupervised text-to-image synthesis [61], and unsupervised image captioning [62], [63], [64], [65], [66]. In particular, unsupervised image cap- tioning has seen flourishing development, driven by two major settings: unpaired image captioning (training with image-text data but unpaired each other) [66], [67], [68] and text-only image captioning (training exclusively on text corpora) [69], [70], [71]. For text-only image captioning, some studies [62], [69], [70] explored leveraging visual concepts to bridge the gap between textual and visual modalities, while others [72], [73], [74] proposed generating synthetic images from text data, training captioning models using these synthetic image-text pairs. Additionally, some research [71], [75], [76] focused on exploiting the cross- modal alignment capabilities of vision-language pretraining models by designing auxiliary modules to better adapt these models to downstream image captioning task. Nevertheless, these advancements in unsupervised vision-language tasks cannot be directly applied to unsupervised TBPS due to the significant differences in tasks. Conventional unsupervised vision-language tasks focus on broad, universal semantic content, often relying on unpaired image-text or text-only corpora. In contrast, unsupervised TBPS emphasizes fine- grained, individual-specific categories using image-only corpora, posing compatibility challenges. 2.3 Image Captioning Image captioning [77] is a vision-language task that aims to generate accurate and fluent natural language descriptions of images. Early approaches [78] relied on simple template- based sentence generation, dependent on object detec- tors or attribute predictors. Subsequent research [79], [80] transitioned to leveraging Convolutional Neural Networks (CNNs) to encode visual features, which were then decoded into coherent textual descriptions using Recurrent Neural Networks (RNNs). Recently, Transformer-based architec- tures [81], [82], [83], [84] have emerged as the dominant approach, being applied both in the visual encoding stage and in the language modeling process. Furthermore, the rapid advancements in multimodal large language models (MLLMs) [85], [86], [87] have introduced new paradigms for image captioning [88], [89], [90]. These models utilized pre- trained vision-language representations to generate more coherent, semantically meaningful captions, significantly advancing the integration of vision and language under- standing. Person images in TBPS differ from image caption- ing, focusing on fine-grained, individual-specific features rather than general semantics. Existing image captioning methods are unsuitable for TBPS. We propose a tired gener- ation framework to enable fine-grained textual descriptions. 2.4 Cross-modal Noise-robust Learning The problem of noisy correspondence in cross-modal learn- ing has garnered increasing attention recently. Huang et al. [91] first identified this issue and introduced a noisy correspondence rectifier that employs neural memorization techniques to categorize data into clean and noisy subsets, subsequently correcting misalignments adaptively. Building on this, Han et al. [92] proposed a meta-learning purification strategy aimed at suppressing noisy samples, Dang et al. [93] developed an information-theoretic framework designed to enhance feature disentanglement in noisy correspondence scenarios. More recently, Shi et al. [94] presented a general framework for filtering out noisy correspondences by utiliz- ing priors from pretrained models. Beyond that, in the TBPS domain, several studies have also addressed the challenge of noisy correspondences. Qin et al. [8] handled misaligned person image-text pairs through a confident consensus mechanism for noise filtering and a triplet alignment loss to avoid misleading supervision during training. Another tackled the issue of noise originating from weakly matched image-text pairs, where parts of the text are irrelevant to the corresponding image, often due to pairing images with generated pseudo-texts. Gao et al. [54] proposed a hybrid patch-channel masking alongside a noise-guided progres- sive training strategy to boost model robustness against noise. Tan et al. [13] introduced a noise-aware masking tech- nique to identify and mask incorrect words based on token similarity. Li et al. [1] further contributed to this field by developing an uncertainty-guided filtration module using a Gaussian mixture model to identify and filter noisy samples. The earlier version of this work [12] explored confidence score-based training to prioritize reliable samples based on fixed confidence scores derived during the generation process. Expanding on this concept, we further develop an adaptive confidence-weighted retrieval learning framework to self-adaptively weight samples. 3 METHOD We focus on solving unsupervised TBPS, in which only a collection of person images S I = I 1 ,I 2 ,· is provided. In the section, we formally delineate the proposed two-stage solution, generation-then-retrieval (as illustrated in Figure 3). The solution consists of a novel tiered description gener- ation framework (detailed in Section 3.1) that automatically generates synthetic text data S T syn = T syn 1 ,T syn 2 ,·, and an adaptive confidence-weighted retrieval learning (de- tailed in Section 3.2) that enables noise-robust training of the retrieval model based on both the original image data S I and the synthetic text data S T syn . For simplicity, we next omit symbolic superscripts and subscripts, using I and T to represent an image and a text, respectively. 3.1 Tiered Description Generation In unsupervised TBPS, there are no annotated text data paired with image data, making the first stage focus on generating pseudo texts for each person image. However, pre-trained vision-language models [11] fall short in cap- turing fine-grained, person-specific details, and template- based methods [1], [12], [14], [56] produce limited, single- IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE5 / Target Image 푇 !"#$% 푞 !"#$% HardNegative Image 푇 &'($ 푞 &'($ MLLM (i) Base Tier: A question-and-answer mechanism MLLM (i) Intermediate Tier: An inter-sample contrastive mechanism ×(ퟏ−휸) Q: What is the person wearing on his top? A: He is wearing a black jacket with ... ... Q: What is he holding in his hand? A: He is holding some paper in his hand. Question& Answer Pairs Q:Whatissimilar...? A:Bothofthemwearbrowpants... Q:Whatisdifferentoftheirtop? A:Theblackjacketwithdark-blueshirt insidewhile...blackdownjacket Question& Answer Pairs Template 푇 )*+, 푇 -./,0 푇 )*+, ... 푇 -./,0 ... ... Expand... ... Diverse Template Gallery ... 푇 '12 푞 '12 (i) Advanced Tier: A stylized expansion mechanism 푇 '12 ×휸 ⨂ 푞 '12 ... ... ∑ 퓛=푪 휷 ·(퓛 푰푻푪 +퓛 푰푻푴 ) ... ... optimize GMM Noisy Sample Clean Sample Visual Encoder Extract:Herearesomehuman description.Extractthedescription templatesfromthesesentences:... 1.Thepersonintheimageis wearing[clothingdescription]... 2.Theperson...a[belongings description]intheirhands. ... Expand:Imitatethesetemplateand providemoredifferentbutusable templates. ... 1.Wearing[clothingdescription] and[footweardescription]... 2.Sporting[hairdescription],the [person/man/woman]is...with [shoesdescription]. ... Extract & Expand Process Extract... Image Gallery ... Extract & Expand Process Retrieval Model A) Generation: tiered description generation B) Retrieval: adaptive confidence-weighted retrieval learning Fig. 3: The overview of the proposed generation-then-retrieval solution (GTR+). The proposed solution consists of a tiered description generation framework and an adaptive confidence-weighted retrieval learning framework. In the first stage, the tiered description generation framework generates fine-grained and diverse textual descriptions through three hierarchical tiers. (i) Base Tier: a question-and-answer mechanism is introduced to generate basic descriptive content; (i) Intermediate Tier: an inter-sample contrastive mechanism enhances descriptions with finer details; (i) Advanced Tier: a stylized expansion mechanism diversifies the linguistic styles of the generated text. In the second stage, the adaptive confidence-weighted retrieval learning framework enables noise-robust training by mitigating the impact of noisy textual descriptions generated in the previous stage. perspective descriptions. To get an enriched description, we propose a three-tiered description generation framework. 3.1.1 Base Tier: Question-and-Answer Mechanism In this tier, we target to generate a basic textual description that captures the essential appearance attributes of the per- son. Rather than the previous version of this work [12] that rely on template-based generation, we leverage an MLLM to generate the basic textual description through a self- question-and-answer approach. Specifically, as illustrated in Figure 4, given an input image, we design a structured instruction prompt to guide the MLLM in generating a set of question-and-answer pairs. These pairs probe explicit attribute information of the person, covering global characteristics (e.g., gender, age) and fine-grained details (e.g., shoe color, accessories). This approach ensures the model performs exhaustive yet grounded attribute queries and responses without requir- ing manual intervention. Subsequently, we apply a sum- instruction to guide the MLLM in extracting and summariz- ing the information from these question-and-answer pairs, thereby generating a basic textual description that naturally captures the key visual attributes of the image. GenerateadescriptionoftheappearanceofthepersonintheimagebasedontheQ&Apairs Iftheanswerisnotsure/clear,justignoreit.Donotimagineanycontentsthatarenotexist. Generatecomprehensivequestionsabouttheappearanceofthepersonin theimage,focusingonbothglobalandfine-grainedattributes(clothing, shoes,hairstyle,gender,age,skintone,belongings,etc.).Askaboutallvisible detailsfromheadtotoe,ensuringnorepeatquestions.Foreachquestion, provideananswerbasedonthevisibleappearanceofthepersonintheimage. Ifanyattributeisnotclearlyvisible,ignoreitanddonotmakeassumptions. 1.Q:Whatisthepersonwearingontheirtop?A:Thepersoniswearingablackjacket. ... 20.Q:Whatisthepersonisholdinginhand?A:Thepersonisholdingsomepaper. Thepersonintheimageisamalewithblackshorthair...Heisalsocarryingablack bagonhisback. Question-and-Answer Pairs Structured Instruction Prompt Sum-Instruction Textual Description (푻 풃풂풔풆 ) Fig. 4: Flowchart of the question-and-answer mechanism. This mechanism enables an end-to-end automated ap- proach for extracting pedestrian attributes and generating textual descriptions, eliminating the need for manual in- tervention and significantly enhancing generalization across diverse images. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE6 GenerateadescriptionoftheappearanceofthepersonintheTarget-ImagebasedontheQ&A pairs.Iftheanswerisnotsure/clear,justignoreit.Donotimagineanycontentsthatarenotexist. IwillprovideTarget-ImageandCompare-Image.Inthesetwoimages,therearetwo individualswhoappearvisuallysimilarbutarenotthesameperson.Yo u rtaskisto comparethevisualfeaturesofthesetwopersonsandgenerateaseriesofQ&Apairs. ... Foreachquestion,provideananswerbasedonthevisibleappearanceofthepersoninthe twoimages.Ifanyattributeisnotclearlyvisible,ignoreitanddonotmakeassumptions. ThepersonintheTarget-Imageis...heiscarryingalargebookorfolderinhislefthand... Q1:Whatarethesimilaritiesintheattireofthepedestrians? A1:Bothpedestriansaredressedin...,thepersoninTarget-Image...andthepersoninCompare-Image... ... Q9:Howdothestuffinthetwopedestrians’handdiffer? A9:ThepedestrianinTarget-Image...whilethepedestrianinCompare-Image... Question-and-Answer Pairs Comparison-oriented Instruction SUM-Instruction Textual Description (푻 풊풏풕풆풓 ) Fig. 5: Flowchart of the inter-sample contrastive mechanism. 3.1.2 Intermediate Tier: Inter-Sample Contrastive Mecha- nism The base tier generates descriptions containing only basic attribute information, and the intermediate tier improves both granularity and informativeness by incorporating hard negative samples. The MLLM is prompted to compare the target image with the hard negative image, emphasizing similarities and differences. This process enriches the de- scription with finer details and deeper contextual informa- tion, improving its quality. Specifically, we first utilize the DINO-v2 [95] to extract the visual feature f img of each image in S I . Next, we compute pairwise similarity between images using cosine similarity, s(f img i ,f img j ) = f img i · f img j ||f img i ||·||f img j || ,(1) where f img i and f img i denote the feature of i-th and j- th (i ̸= j) images, repsectively. For the target image, we identify its top-10 nearest neighbors based on pairwise similarity and randomly select one as a hard negative image. Instead of always selecting the top-1 nearest neighbor as the hard negative, the random selection from the top-10 nearest neighbors ensures greater diversity in the negative samples. The picked hard negative, while visually similar to the target, is semantically distinct, providing a challenging yet informative contrastive example. Given the target image and its paired hard negative image, as shown in Figure 5, we enhance the instruction prompt from the base tier by introducing a comparison- oriented instruction. This enhanced instruction guides the MLLM to generate a set of question-and-answer pairs that highlight both shared visual elements (e.g., ’both subjects wear striped shirts’) and distinctive visual features (e.g., ’distinct shoe colors: white vs. black’) between two images. Subsequently, we apply a revised SUM-Instruction, specifi- cally tailored for the comparison phase, to the MLLM for generating the final text description. This revised SUM- Instruction ensures that the generated caption remains fo- cused solely on the target image while being grounded in the comparative context provided by the hard negative image. 3.1.3 Advanced Tier: Stylized Expansion Mechanism Through the aforementioned two tiers, we obtain textual descriptions that capture essential appearance attributes and fine-grained details of the person images. However, we observe that MLLMs tend to generate captions with limited sentence structures when provided with identical instruction prompts. This structural uniformity in the de- scriptions could result in overfitting to specific linguistic patterns, thereby diminishing the model’s ability to gen- eralize effectively in real-world applications. To address this, we consider to explicitly diversify the styles of textual descriptions in this tier. Specifically, we begin by randomly sampling a subset of descriptions obtained from the aforementioned two tiers and using ChatGPT-4o to analyze their underlying syntactic templates. Through multiple rounds of interaction with ChatGPT-4o, we iteratively refine these templates to create a diverse array of sentence patterns, which are incorporated into a diverse template gallery. During textual description generation, a template is randomly selected from the diverse template gallery. The MLLM is then prompted to follow the chosen template and generate a detailed textual description (denoted as T adv ) using the following instruction, Generate a detailed description of the person’s overall appearance based on the T base and T inter , using the following template:template. Do not invent or assume any details that are not evident in the image. Here, T base and T inter denote the texts generated by the base tier and the intermediate tier, respectively, and template is placeholder that is replaced with the chosen template. Finally, for each image, T adv is generated as the paired text. Since T adv already incorporates descriptive information from both T base and T inter , we use only T adv in constructing the final image-text pair. To further enhance the diversity of the generated texts, we leverage two state-of-the-art, publicly available MLLMs: Qwen2-VL-7B-Instruct [96] and InternVL2.5-8B [97] for generation, respectively. As a result, each image is paired with two texts, forming the final training dataset. 3.2 Adaptive Confidence-Weighted Retrieval Learning The generative stage produces synthetic image-text pairs that, in principle, could serve as supervised signals for train- ing the retrieval model. However, the inherent propensity of MLLMs to introduce hallucinations and inaccuracies during generation leads to noise, thus hindering downstream re- trieval performance. To mitigate the effects of such noise, we develop an adaptive confidence-weighted retrieval learning framework. 3.2.1 Baseline We first adopt BLIP [11] as the baseline retrieval model, and yet it can be replaced with others (as verified experimentally in Section 4.6.2) due to the flexible nature of the proposed retrieval learning framework. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE7 BLIP consists of an image encoderE img , a text en- coderE txt , and an image-grounded text encoderE i&t . Given an image I = p 1 ,p 2 ,· and a generated text T = w 1 ,w 2 ,·, where p i is i-th non-overlapping patch and w j is j-th token, we compute the global feature of image, the global feature of text, and the multimodal feature of the image-text pair as, f img =E img (I), f txt =E txt (T),(2) f i&t =E i&t (E img (I),E txt (T)).(3) Then, the optimization objectives with an image-text contrastive learning (ITC) loss and an image-text matching (ITM) loss are computed as, L i2t itc =−E (I,T)∼D log exp(s(f img ,f txt )/τ) P M m=1 exp(s(f img ,f txt m )/τ) , L t2i itc =−E (I,T)∼D log exp(s(f img ,f txt )/τ) P M m=1 exp(s(f img m ,f txt )/τ) , L itc = (L i2t itc + L t2i itc ) / 2, (4) where M is the number of instances in a mini-batch, s(f img ,f txt ) measures the cosine similarity between the image and the text, τ is a learnable temperature; L itm = E (I,T)∼D H(y,φ(f i&t )),(5) where H denotes the cross-entropy function, and y is 2- dimension one-hot vector and denotes the ground-truth label. φ(f i&t ) is the predicted matching probability of the pair and φ(·) is a linear classification layer. 3.2.2 Adaptive Confidence Weighting for Noise-Robust Re- trieval In the baseline retrieval model, the training objectives (Eq. 4 and Eq. 5) treat all image-text pairs equally, assigning uni- form weights during optimization. However, noise in the generated texts can mislead training and degrade perfor- mance. To address this, we introduce adaptive confidence weighting into the objectives, assigning higher weights to clean samples and lower weights to noisy ones, thereby mitigating the impact of noise. The confidence score reflects the likelihood of the pair being clean, with high value indicating clean data and vice versa. To estimate the score, we model the similarity dis- tribution of image-text pairs in each training epoch. Based on this distribution, we compute the probability that each pair belongs to the clean data distribution, which is then used as the confidence score for weighting in the subsequent training step. Specifically, we employ the Gaussian Mixture Model (GMM), which is well known for its ability to capture com- plex distributions by combining multiple Gaussian com- ponents, to model the similarity distribution of image-text pairs, p(ˆs) = K X k=1 α k N(ˆs| μ k ,σ 2 k ).(6) Here, we set K = 2, corresponding to two underlying distributions: one representing clean data and the other representing noisy data. The parameter α k denotes the mixing weight of the k-th Gaussian component, satisfying α 1 +α 2 = 1 and α k ≥ 0 for all k. The functionN(ˆs| μ k ,σ 2 k ) denotes the probability density function of a univariate Gaussian, where μ k and σ 2 k are the mean and variance of the k-th component, respectively. And ˆs represents the positive pair similarity and is computed as, ˆs(f img ,f txt ) = γ· s(f img ,f txt ) + (1− γ)· q txt ,(7) where the cosine similarity s(f img ,f txt ) is rectified by in- tegrating the text generation probability q txt in γ-weighted manner. Specifically, during the previous tiered description generation stage, token-level generation probabilities can outputted from the MLLM. Based on them, the text gen- eration probability q txt is calculated as the mathematical expectation of the generation probabilities of its constituent tokens, serving as a measure of central tendency: q txt = N X i 1 N q tok i ,(8) where q tok i is i-th token generation probability and N is the number of tokens composing the text. A lower value of q txt indicates higher uncertainty in the correspondence between the generated text and the associated image, im- plying potential noise. Notably, the previous version of this work [12] computed the confidence score as the joint proba- bility of token-level scores, which often results in vanishing values for longer captions and thus limited effectiveness. In contrast, our expectation-based confidence measure offers a more objective and robust assessment of text confidence. The Gaussian Mixture Model in Eq. 6 is optimized using the Expectation-Maximization (EM) algorithm, based on which we compute the posterior probability that a positive pair belongs to the clean data distribution as its confidence score, C = α k ∗ N(ˆs| μ k ∗ ,σ 2 k ∗ ) P 2 k=1 α k N(ˆs| μ k ,σ 2 k ) , k ∗ = arg max k μ k .(9) Here, the component with the larger mean, μ k ∗ , represents the clean data distribution. Finally, we incorporate the confidence score into the optimization objectives as, L i2t itc =−E (I,T)∼D C β log exp(s(f img ,f txt )/τ) P M m=1 exp(s(f img ,f txt m )/τ) , L t2i itc =−E (I,T)∼D C β log exp(s(f img ,f txt )/τ) P M m=1 exp(s(f img m ,f txt )/τ) , L itc = (L i2t itc + L t2i itc ) / 2, (10) L itm = E (I,T)∼D C β H(y,φ( ˆ f i&t )),(11) where β is a hyper-parameter to control the importance of the confidence score. During each training batch, the model with confidence- weighted objectives updates the similarity distributions and returns the dynamically updated parameters of the GMM, allowing confidence scores to adapt accordingly. These scores are then incorporated into the training objectives. This alternating iterative process ensures that confidence es- timation remains sensitive to the current training dynamics, yielding more accurate assessments of data noise compared IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE8 LargeFine-Person Dataset Themaniswearingablueandredbaseballcap,awhiteshirtwithasmallbluelogo,lightbrownpants, anddarkshoes.Theyhavealightskintoneandablackwristbandontheirrightwrist.Theperson's hairstyleisnotclearlyvisible.Therearenovisiblebelongingsoritemsthatthepersonisholding. Advanced Tier: Stylized Expansion Mechanism ThepersoninTarget-Imageiswearingawhitet-shirtpairedwithasmallbluelogo,light-coloredpants, anddarkshoes.Theyhaveashort,neathairstyleandappeartobemale.Theirageisestimatedtobein theearlytwenties,andtheyhavealightskintone.Theyarenotcarryinganyvisibleitems. Intermediate Tier: Inter-Sample Contrastive Mechanism Thepersonintheimageiswearingablueandredbaseballcap,awhiteshirt,lightbrownpants,and darkshoes.Theyhavealightskintoneandablackwristbandontheirrightwrist.Theperson's hairstyleisnotclearlyvisible,andtherearenodistinguishablefeaturesontheirpants.Thebackground appearstobeawhiteandyellowindoorsetting,possiblyastaircasearea.Therearenovisible belongingsoritemsthatthepersonisholding. Base Tier: Question-and-Answer Mechanism Dressedinasleevelessyellowtopwithabluesectionontherightarmandaredskirtthatappearsto beashort,A-linestyle,thewomaniswearingwhiteshoes.Shehasblackhairtiedbackinaponytail andisholdingabatoninherhands. Advanced Tier: Stylized Expansion Mechanism ThepersoninTarget-Imageiswearingasleevelessyellowtopwithabluesectionontherightarm, pairedwitharedskirtandwhiteshoes.Theirhairisstyledinaponytail.Theyappeartobemale,and theirskintoneislightandfair.Theindividualisholdingabaton. Intermediate Tier: Inter-Sample Contrastive Mechanism Thepersonintheimageisayoungfemalewithfairskinandblackhairtiedbackinaponytail.Sheis wearingasleevelessyellowtopandaredskirtthatappearstobeashort,A-linestyle.Sheisalso wearingwhiteshoesandisholdingamicrophoneinherhands.Therearenoadditionallayersor clothingitemsarebeingworn.Theoverallcolorofthetopisyellow,whiletheskirtisred. Base Tier: Question-and-Answer Mechanism Noise images Low-resolution images Blurred images Abnormal number of individuals Fig. 6: Illustration of the proposed LargeFine-Person dataset and the screened noise images during dataset construction. To ensure data diversity and scale, we generate three textual descriptions for each image, corresponding to the three tiers of our tiered description generation framework, which capture different aspects of person appearance and attributes. to using a fixed text generation probability q txt as the con- fidence score—as done in prior work [12]. Moreover, due to Eq. 7, q txt provides relatively reliable guidance in the early stages of training, when the model is still unstable and the cosine similarity s(f img ,f txt ) may be unreliable. As train- ing progresses and features become more discriminative, the cosine similarity increasingly captures meaningful semantic relationships. By jointly modeling both signals within a uni- fied framework, the β-weighted addition ensures robustness to noisy data across different training phases. 3.3 Pre-training Dataset Construction: LargeFine- Person In addition, leveraging the proposed tiered description generation framework, we introduce LargeFine-Person, a large-scale and fine-grained TBPS dataset. It is specifically designed to serve as a generalizable pre-training benchmark for TBPS in unsupervised setting. Image Collection. We adopt LUPerson [17] and LPW [18] datasets as the image sources. LUPerson contains 4 million (M) person images cropped from YouTube videos, while LPW includes 0.59M person images captured by multi-camera systems. Together, these datasets provide a rich variety of environments. However, since the images in LUPerson are extracted from videos using algorithms, they inevitably contain a substantial amount of noise. Thus, we perform data denoising on LUPerson to filter out low- quality images based on the following criteria: •Low-resolution images: Images with a width smaller than 32 pixels or a height smaller than 64 pixels are filtered out. •Blurred images: We use the variance of the Laplacian response to measure image sharpness. Images with a Laplacian variance smaller than 80 are regarded as blurry and filtered out. •Abnormal person instances: We utilize YOLOv8 for person detection. Images with no detection box or more than one detection box are filtered out. As a result, we obtain 0.61M high-quality images from the LUPerson dataset, which, combined with the 0.59M images from the LPW dataset, yield a total of 1.2M images. Text Generation. Using the tiered description generation framework, we generate textual descriptions for each of the 1.2M images, incorporating all generated texts (T base , T inter , and T adv ) to enhance diversity. To maintain gener- ation efficiency, we randomly utilize either Qwen2-VL-7B- Instruct [96] or InternVL2.5-8B [97] within this framework. Finally, this process yields three textual descriptions per image, resulting in a total of 3.6M image-text pairs. In summary, the proposed tiered description generation framework enables the construction of LargeFine-Person entirely without manual annotation. The characteristics of LargeFine-Person are summarized in Table 1, and example illustrations are shown in Figure 6. 4 EXPERIMENT 4.1 Datasets Our experiments involve three open benchmark datasets and one proposed pre-training dataset. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE9 CUHK-PEDESICFG-PEDESRSTPReid MethodsReference R@1R@5R@10mAPR@1R@5R@10mAPR@1R@5R@10mAP Supervised TBPS Dual Path [23]TOMM-2044.4066.2675.07-38.9959.4468.41----- ViTAA [30]ECCV-2055.9775.8483.5251.60 50.9868.7975.78----- CMKA [98]TIP-2054.6973.6581.86-46.4967.1475.18----- LapsCore [40]ICCV-2163.40-87.80--------- SRCF [26]ECCV-2264.0482.9988.81- 57.1875.0181.49----- IRRA [99]CVPR-2373.3889.9393.7166.1363.4680.2585.8238.0660.2081.3088.2047.17 RaSa [2]IJCAI-2376.5190.2994.2569.3865.2880.4085.1241.2966.9086.5091.3552.31 MACF [100]IJCV-2473.3388.5793.02- 62.9579.9385.04----- TBPS-CLIP [45]AAAI-2473.5488.1992.3565.3865.0580.3485.4739.8361.9583.5588.7548.26 AUL ∗ [16]AAAI-2477.2390.4394.41- 69.1683.3288.37-71.6587.5592.05- RDE [8]CVPR-2475.9490.1494.1267.5667.6882.4787.3640.0665.3583.9589.9050.88 PTMI [38]TIFS-2576.0289.9394.1470.8566.5481.5086.5046.0563.3581.9589.1052.46 GAHR [28]TIFS-2576.6490.4694.1066.81 68.6982.8387.4042.1068.8586.5091.1053.60 ICL [47]CVPR-2578.1891.6394.8369.5869.2283.4988.0642.3470.0086.6091.7054.16 Weakly-supervised TBPS CMMT [49]ICCV-2157.1078.1485.23--------- CPCL [50]arXiv-2470.0387.2891.7863.1962.6079.0984.4636.1658.3581.0587.6545.81 DG-CMIA [53]arXiv-2573.0689.2193.4464.8863.7179.9085.3836.8761.3082.0088.1047.20 Semi-supervised TBPS STBPS [54]arXiv-2463.8782.2087.7057.1846.4664.3471.6026.9056.4578.9587.0544.45 Unsupervised TBPS GTR [12]ACMMM-2347.5368.2375.9142.9128.2545.2153.5113.8245.6070.3579.9533.30 GAAP [56]IJCAI-2447.6467.7976.0841.2827.1244.9153.5611.4344.4565.1575.3031.21 MUMA [1]AAAI-2559.5277.7984.6552.7538.1156.0163.9619.0254.3576.0583.6540.50 GTR+Ours61.3579.3585.7555.7547.8164.9771.9428.7554.7575.1583.5043.79 GTR+(Pre-training)Ours64.6580.7286.7858.6752.7867.9473.9133.9955.7076.5584.2543.86 TABLE 2: Performance comparison of TBPS methods under different settings. Supervised TBPS uses a large number of images with manually annotated texts for training. Weakly-supervised TBPS only requires image-text pairs, without identity annotations. Semi-supervised TBPS leverages a small set of annotated images (1% of all training data) and a large set of unannotated ones. Unsupervised TBPS operates without any textual annotations. GTR + (Pre-training) denotes that our proposed GTR + can be pre-trained on our LargeFine-Person dataset. CUHK-PEDES [22] is the most commonly-used dataset in TBPS. It consists of 40, 206 images and 80, 440 texts from 13, 003 identities in total, which are split into 34, 054 images and 68, 126 texts from 11, 003 identities in the training set, 3, 078 images and 6, 158 texts from 1, 000 identities in the validation set, and 3, 074 images and 6, 156 texts from 1, 000 identities in the test set. The average length of all texts is 23. ICFG-PEDES [101] contains 54, 522 images from 4, 102 identities in total. Each of the images is described by one text. The dataset is split into 34, 674 images from 3, 102 identities in the training set, and 19, 848 images from 1, 000 identities in the test set. On average, there are 37 words for each text. RSTPReid [102] consists of 20, 505 images of 4, 101 iden- tities. Each identity has 5 corresponding images captured from different cameras. Each image is annotated with 2 textual descriptions, and each description is no shorter than 23 words. There are 3, 701/200/200 identities utilized for training/validation/testing, respectively. LargeFine-Person is our proposed large-scale, fine- grained description dataset for pre-training TBPS. It com- prises 1.2M images covering 611, 460 identities, with each image annotated with 3 textual descriptions on average. The descriptions exhibit rich linguistic variation and fine- grained detail, with an average length of 61 words per description, and up to 507 words in total. During the experiments, unless otherwise specified, we evaluate the proposed GTR+ on each dataset under an unsupervised setting. Specifically, only the images from the training set are used for training, while the corresponding textual annotations in the training set are excluded during the training process. 4.2 Protocol We adopt the widely-used Rank@K (R@K for short, K=1, 5, 10) metric to evaluate the performance of the proposed method. Specifically, given a query text, we rank all the test images via the similarity with the query text, and the search is deemed to be successful if top-K images contain any corresponding identity. R@K is the percentage of successful searches. In addition, we also adopt the mean average pre- cision (mAP) as a complementary metric. Rank@K reflects the accuracy of the first few retrieval results, while mAP emphasizes the comprehensive performance of the method. 4.3 Implementation Details We train the retrieval model using image inputs resized to 256× 256 and textual inputs truncated to a maximum of 77 tokens. The model is optimized using the Adam optimizer with a learning rate of 1e-6, a batch size of 25, and trained for 15 epochs. All experiments are conducted on NVIDIA IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE10 CUHK-PEDESICFG-PEDESRSTPReid Pre-training DatasetsBaselinesFine-tuning R@1R@5R@10mAPR@1R@5R@10mAPR@1R@5R@10mAP ×20.3238.0948.3819.149.1020.6428.243.8622.2044.5558.1517.03 RDE [8] ✓53.5173.8081.6948.23 42.4458.9866.3321.8851.8072.8080.9538.75 ×21.1038.5448.5219.5110.6721.9528.194.7322.5145.6859.3617.11 IRRA [99] ✓57.6575.8182.3451.46 41.1259.1367.0721.9649.1072.8581.4537.52 ×26.9846.1855.3623.5912.4722.0428.225.3026.1552.2064.2019.15 LUPerson-T [19] BLIP [11] ✓61.1679.5586.3555.75 45.8463.3070.1526.6551.3074.5081.8541.02 ×54.3973.0280.3448.4836.4154.9963.4918.8449.8072.4582.0037.05 RDE [8] ✓55.8876.3282.9449.7342.8060.2067.6122.8852.8073.8083.0039.29 ×57.6075.9482.7751.4438.3056.6064.5420.4351.5073.9282.6537.34 IRRA [99] ✓60.4179.5685.5154.2446.0062.4569.6324.6754.7075.4084.1040.14 ×60.1177.4584.1553.4641.9759.7567.0120.8152.4075.1582.8038.81 LUPerson-MLLM [13] BLIP [11] ✓64.5280.0786.9157.7449.8966.4073.1230.7354.9076.2083.8544.67 ×56.4575.0582.1050.0541.0559.1566.7520.7850.7073.6582.4037.15 RDE [8] ✓56.9775.9382.7850.7844.4760.7167.7523.7850.8572.1581.7038.42 ×59.4478.5485.2254.1143.7760.7768.0522.3050.4573.4582.3537.68 IRRA [99] ✓60.8880.0286.4555.5446.0262.7669.9224.5854.4076.0584.1041.31 ×62.6578.8084.7655.2747.5364.3271.3925.3852.0074.0582.3538.72 LargeFine-Person (Ours) BLIP [11] ✓64.6580.7286.7858.6752.7867.9473.9133.9955.7076.5584.2543.86 TABLE 3: Comparison of pre-training datasets under the unsupervised setting. Only two pre-training datasets, LUPerson-T and LUPerson-MLLM, conform to the unsupervised setting (see Table 1) and are used for a fair comparison.×Fine-tuning: the baseline is only trained on the pre-training dataset and evaluated directly.✓Fine-tuning: the baseline is further fine- tuned on the respective downstream dataset (images from the original dataset paired with synthesized descriptions via our tiered description generation framework). Best and second-best results are denoted in bold and underline, respectively. L40 GPUs. In addition, the hyper-parameter γ in Eq. 7 is set to 0.2, and β in Eq. 10 and Eq. 11 is set to 0.8. 4.4 Comparisons with the State-of-the-Art Methods We compare the proposed GTR+ with SOTA methods on CUHK-PEDES, ICFG-PEDES and RSTPReID in Ta- ble 2. These SOTA methods fall into four distinct learn- ing paradigms. Supervised TBPS relies on fully annotated image-text pairs; weakly supervised TBPS requires image- text pairs without identity labels; semi-supervised TBPS leverages a small annotated set (i.e., 1% of data) and a large amount of unannotated images; unsupervised TBPS operates entirely without textual annotations. Among these settings, the unsupervised paradigm provides the most fair and direct comparison with our GTR+. In contrast, the other paradigms leverage additional supervision sig- nals, which naturally confer a performance advantage. As expected, GTR+ exhibits a performance gap compared to methods that utilize more supervision signals, due to their access to richer annotated data. Whereas, compared to unsupervised methods, GTR+ achieves significantly supe- rior performance, demonstrating its effectiveness. Notably, when further pre-trained on our constructed LargeFine- Person dataset—still under an unsupervised setting—GTR+ achieves even stronger results, even outperforming semi- supervised method. This highlights the potential of our constructed large-scale dataset. Furthermore, we compare the proposed LargeFine- Person pre-training dataset with other pre-training datasets in terms of performance, To this end, we pre-train various retrieval baselines on each dataset, and the results are pre- sented in Table 3 and Table 4. ❶ We first conduct compar- isons under unsupervised setting. Notably, only LUPerson- T and LUPerson-MLLM are constructed independently of human annotations, thus aligning with the unsupervised paradigm. For a fair comparison, we compare our dataset Pre-training DatasetsR@1R@5R@10mAP CUHK-PEDES w/ Human-annotation MALS [20]74.0589.4893.6466.57 SYNTH-PEDES [14]74.8889.5894.2967.10 HAM-PEDES [15]77.7191.4294.5769.68 w/o Human-annotation LUPerson-T [19]74.3789.5193.9766.60 LUPerson-MLLM [13]76.8291.1694.4669.55 LargeFine-Person (Ours)77.1390.8294.4968.37 ICFG-PEDES w/ Human-annotation MALS [20]64.3780.7586.1238.85 SYNTH-PEDES [14]64.3180.9286.5338.49 HAM-PEDES [15]68.2583.3088.1542.30 w/o Human-annotation LUPerson-T [19]64.5080.2485.7438.22 LUPerson-MLLM [13]67.0582.1687.3341.51 LargeFine-Person (Ours)67.8082.8187.6641.00 RSTPReid w/ Human-annotation MALS [20]61.9080.6089.3048.08 SYNTH-PEDES [14]64.0082.8589.9050.06 HAM-PEDES [15]71.6987.8593.3055.19 w/o Human-annotation LUPerson-T [19]62.2083.3089.7548.33 LUPerson-MLLM [13]68.5087.1592.1053.02 LargeFine-Person (Ours)69.0586.9092.2554.19 TABLE 4: Comparison of pre-training datasets under super- vised setting. We adopt IRRA [99] as the retrieval baseline. against these two pre-training datasets. As shown in Table 3, when the retrieval baseline is trained on the pre-training dataset and evaluated directly (i.e., without fine-tuning), our LargeFine-Person consistently achieves superior perfor- mance across all baselines, demonstrating its effectiveness. Furthermore, owing to our proposed tiered description gen- eration framework, downstream datasets—CUHK-PEDES, ICFG-PEDES, and RSTPReid—can be expanded by synthe- sizing paired textual descriptions using only image data, fully adhering to the unsupervised setting. The baseline models are then fine-tuned on these synthesized datasets IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE11 Base tierIntermediate tierAdvanced tier Methods Temp.Q&ARand.ISCIntg.StyExp R@1R@5R@10mAP TierGBase T ✓47.2767.5875.6242.28 TierG Base Q ✓54.8273.9180.9949.42 TierGInterm R ✓53.3373.1680.4947.80 TierGInterm I ✓54.5873.8581.0349.31 TierG I ✓55.2374.2481.5149.72 TierG✓55.5874.4081.8750.12 TABLE 5: Ablation Study of the proposed tiered description generation framework (TierG) on CUHK-PEDES. Q&A: question-and-answer mechanism; ISC: inter-sample contrastive mechanism; StyExp: stylized expansion mechanism. Temp.: template-based generation (vs. Q&A); Rand.: random negative sampling (vs. ISC); Intg.: direct integration without stylized expansion (vs. StyExp). R@1R@5R@10mAP Baseline55.5874.4081.8750.12 k=157.9476.9283.7751.66 k=558.4577.2584.0152.01 k=10 (Ours)59.5078.4984.6753.56 TABLE 6: Sensitivity analysis of the top-k choice for hard negative selection in the intermediate tier on CUHK-PEDES. Best results are denoted in bold. (i.e., with fine-tuning). It can be seen that LargeFine-Person continues to outperform the competing datasets across all baselines in this scenario as well. Moreover, we observe a significant performance gain across all baselines after fine- tuning, which further validates the high-quality of the syn- thesized downstream dataset and the effectiveness of our tiered description generation framework. ❷ We further con- duct comparisons with other pre-training datasets under the supervised setting. In this setup, the retrieval model is first pre-trained on the pre-training dataset and then conven- tionally fine-tuned on the downstream dataset. As shown in Table 4, the model pre-trained on our LargeFine-Person achieves competitive performance compared to other un- supervised pre-training datasets (i.e., those constructed without human annotation). However, it is slightly infe- rior to pre-training datasets that leverage human-annotated data during construction—particularly on the HAM-PEDES dataset. This performance gap is reasonable: HAM-PEDES is generated with reference to human-annotated textual descriptions, which ensures higher semantic fidelity. In con- trast, LargeFine-Person is constructed entirely without hu- man supervision. As a result, HAM-PEDES contains higher- quality image-text pairs, leading to stronger pre-training efficacy. 4.5 Ablation Study 4.5.1 Effectiveness of the tiered description generation framework We propose a tiered description generation framework (TierG) for textual description generation, which incorpo- rates three key components: the base tier with the question- and-answer mechanism (Q&A), the intermediate tier with the inter-sample contrastive mechanism (ISC), and the ad- vanced tier with the stylized expansion mechanism (Sty- Exp). In this part, we conduct an ablation study to evaluate the contribution of each component, as summarized in Ta- ble 5. ❶ The goal of the base tier is to generate a basic textual description. In previous work [12], we employ a template- based approach (Temp.), in which the instruction is used to prompt the vision-language model to extract attributes, followed by filling these attributes into a manually designed template to construct the final description. This serves as a comparison for our proposed Q&A mechanism. Compared to Temp., Q&A offers clear performance advantages. Since Q&A provides greater flexibility in attribute generation and eliminates the need for laborious manual template design, enabling more natural and varied descriptions. ❷ For the intermediate tier, the objective is to generate fine-grained textual descriptions. For this, we introduce hard negative images selected from the top-k most similar images as comparison during the generation process (ISC). In con- trast, randomly selecting a negative image from all images (Rand.) is a feasible solution, but leads to the degradation in performance. The use of hard negatives in ISC provides more informative contrastive signals, which significantly benefits the model in generating more precise and dis- criminative descriptions. ❸ In the advanced tier, we aim to generate descriptions with diversified styles. For this, we design an instruction that builds upon the outputs from the first two tiers, guided by an explicitly stylized template to steer the MLLM toward generating stylistically rich text (StyExp). Alternatively, we can omit the stylized template and simply integrate the descriptions from the base and intermediate tiers for generation (Intg.). Results show that StyExp achieves superior performance, demonstrating that explicitly diversifying the style of textual descriptions is effective for improving TBPS performance. ❹ Overall, our proposed TierG employs a three-tier sequential archi- tecture to generate high-quality textual descriptions in a progressive manner. Results show that TierG outperforms TierG Base Q and TierGInterm I , which strongly validates the effectiveness of the tiered design. 4.5.2 Effectiveness of the adaptive confidence-weighted retrieval learning framework We propose an adaptive confidence-weighted retrieval learning framework (AdaWeight), to mitigate the impact of noisy data introduced during the text generation stage. To evaluate the effectiveness of AdaWeight, we compare with several alternative strategies. Tan et al. [13] introduced LUPerson-MLLM, a large-scale TBPS dataset constructed IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE12 MaskingWeighting Methods Text (sim.)Text (prob.)ImageFixed conf.Adaptive conf. R@1R@5R@10mAP Baseline55.5874.4081.8750.12 Sim-TxtMask [13]✓55.6074.4581.9349.26 Prob-TxtMask✓55.6774.5182.1350.76 Rand-ImgMask✓56.6875.6282.3451.26 FixWeight [12]✓58.7277.7184.2854.10 AdaWeight✓61.3579.3585.7555.75 TABLE 7: Ablation Study of the proposed adaptive confidence-weighted retrieval learning framework on CUHK-PEDES. Noise-robust learning can be achieved through two main strategies. (1) Masking noisy tokens/patches, including: masking text tokens with low image-text similarity scores (Text (sim.)), masking text tokens with low generation probabilities (Text (prob.)), randomly masking image patches (Image); (2) Weighting training samples based on noise levels, including: Fixed confidence weighting (Fixed conf.), Adaptive confidence weighting (Adaptive conf.). R@1R@5R@10mAP Baseline55.5874.4081.8750.12 K=157.2977.2383.5651.21 K=2 (Ours)61.3579.3585.7555.75 K=359.3178.4684.8453.54 K=459.4978.4084.6853.79 K=559.3478.6284.7853.61 TABLE 8: Sensitivity analysis of the number of Gaus- sian components in GMM-based confidence estimation on CUHK-PEDES. K denotes the number of Gaussian compo- nents. Best results are denoted in bold. using synthetic textual descriptions, and proposed a noise- robust retrieval method by masking text tokens with low image-text similarity scores. Inspired by this work, we observe that besides the weighting strategy, the masking strategy can also serve as an effective means for noise- robust learning. We present the comparison results in Ta- ble 7. ❶ Masking strategy for noise-robust learning. We evaluate three masking approaches: (i) Following [13], we mask text tokens with low similarity scores relative to the paired images (Sim-TxtMask); (i) Alternatively, we mask tokens with low generation probabilities from the gener- ation model (Prob-TxtMask); (i) We extend this idea to the image domain by randomly masking image patches (Rand-ImgMask), which can implicitly reduce the influence of noisy correspondences. From Table 7, Rand-ImgMask achieves the best performance, showing a slight improve- ment over text-based masking variants. The image typically exhibits more semantic coherence than the text and retains more meaningful information after partial masking, en- abling more robust cross-modal alignment under noisy con- ditions. ❷ Weighting strategy for noise-robust learning. In previous work [12], a fixed confidence-weighting approach (FixWeight) is proposed based on text generation probabil- ities: higher probabilities indicate more confident genera- tions, implying less noise, and thus greater weights. How- ever, FixWeight fails to capture dynamic changes during training, potentially leading to suboptimal handling of noisy samples. In contrast, our proposed adaptive confidence- weighted approach (AdaWeight) dynamically adjusts sam- ple importance according to both generation confidence and training progress, offering a more principled and flexible solution. ❸ Masking vs. Weighting. Overall, the weighting strategy demonstrates superior effectiveness compared to the masking strategy, with our proposed AdaWeight achiev- ing the best performance and offering clear advantages. 4.5.3 Hyper-parameter analysis ❶ Top-k hard negative selection. We evaluate the sensitivity of the top-k hard negative selection during the intermediate generation tier, as shown in Table. 6. While all settings outperform the baseline, k = 10 yields the best results. An overly small k (e.g., k = 1) restricts candidates to a narrow neighborhood, limiting comparison diversity. Con- versely, sampling from the top-10 neighbors strikes an op- timal balance: it preserves visual similarity while providing diverse contrastive cues, which generates richer discrimi- native descriptions and improves retrieval performance. ❷ Gaussian components (K) and GMM discriminative ca- pability. To explicitly validate the discriminative capability of our GMM, Figure. 7 visualizes its stage-wise behavior. The model consistently separates training pairs into clean and noisy subsets. More importantly, the similarity gap between these two subsets gradually widens as training proceeds. This dynamic confirms that the GMM is not merely fitting a static bimodal shape, but progressively and reliably identifying clean samples as the feature representa- (a) Epoch 1 Density Raw image-text similarity Gap (c) Gap (b) Epoch 15 Raw image-text similarity Gap=2.02 0.14 0.12 0.10 0.08 Density 0.14 0.12 0.10 0.08 0.06 0.04 0.02 0.00 0 5 1015202530 Assigned Noisy Assigned Clean Clean-Noisy Gap 2.6 2.4 2.2 2.0 2.8 0.06 0.04 0.02 0.00 Epoch 123456789101112131415 0 5 1015202530 Gap=2.83 Fig. 7: Visualization of GMM-based separation during train- ing. Panels (a) and (b) show the distributions of image- text similarity for the subsets assigned by the GMM to the noisy and clean components at epochs 1 and 15, respectively. Panel (c) shows the gap between the clean and noisy subsets over training epochs. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE13 50 55 60 65 00.20.40.60.81 R@1 mAP 45 50 55 60 65 00.20.40.60.81 mAP R@1 휸휷 Fig. 8: Hyper-parameter analysis of γ and β. tions improve. Furthermore, we justify our two-component design by analyzing the sensitivity of the Gaussian com- ponent number (K ), as reported in Table. 8. The results demonstrate that K = 2 yields the optimal performance. A single component (K = 1) is too coarse to model the underlying noise variation, whereas using more compo- nents (K > 2) tends to over-partition the distribution, leading to degraded confidence estimation. ❸ Impact of γ and β. Our adaptive confidence-weighted retrieval learning framework involves two hyper-parameters: γ and β. The parameter γ balances the contributions of pair-wise cosine similarity and text generation probability in computing the confidence scores, while β controls the influence of these confidence scores within the training objectives. A higher γ places more weight on cosine similarity, whereas a lower γ emphasizes text generation probability. And γ = 1 corre- sponds to adopting only cosine similarity for computing the confidence scores, and γ = 0 corresponds to using only text generation probability. As shown in Figure 8, relying solely on either component (γ = 0 or γ = 1) leads to suboptimal performance. The best results are achieved by combining both with γ = 0.2. The cosine similarity is an optimizable dynamic value that improves during training and a smaller γ helps maintain training stability. The parameter β governs the overall impact of confidence scores on training. When β = 0, confidence scores have no effect. As illustrated in Figure 8, performance is relatively low when β < 0.5, becomes more stable for β > 0.5, and reaches its peak at β = 0.8. 4.6 Generalization Experiment 4.6.1 Domain generalization We evaluate the domain generalization capability of the proposed adaptive confidence-weighted learning frame- work in Table 9. Specifically, models are trained on the source dataset and directly tested on the target dataset, e.g., C→I, where training is conducted on CUHK-PEDES and evaluation is performed on ICFG-PEDES. Notably, all compared methods are supervised approaches that rely on both images and human-annotated textual descriptions during training. Moreover, among these, MRA [103] is TBPS method specifically designed for the domain generalization setting. In contrast, our GTR+ operates in an unsupervised manner, requiring only images from the source domain, with textual descriptions automatically generated by our proposed tiered description generation framework. Despite the absence of annotated texts, GTR+ achieves competitive performance, demonstrating the high quality of the syn- MethodsR@1R@5R@10mAP C → I LGUR [36]34.2552.2860.85- IRRA [99]42.6262.3970.1323.89 BLIP [11]47.9363.9871.0726.25 RDE [8]48.1566.3673.4327.01 GAHR [28]49.0066.3473.59- MRA* [103]50.0167.1374.24- GTR+ (Ours)48.9365.9871.2728.25 I → C LGUR [36]25.4444.4854.39- IRRA [99]32.2155.1265.1931.72 BLIP [11]36.3557.9867.2134.19 RDE [8]38.3359.5168.3035.56 GAHR [28]40.6062.4671.62- MRA* [103]47.6668.7076.82- GTR+ (Ours)38.7360.1269.8735.90 TABLE 9: Domain generalization comparisons. C: CUHK- PEDES, I: ICFG-PEDES. GTR+ trains in source with an unsupervised setting with generated texts, others rely on fully annotated supervised data. * denotes the method specifically designed for the domain generalization scenario. BaselinesR@1R@5R@10mAP RDE [8]59.9378.8984.1853.69 IRRA [99]60.7279.1184.9354.80 BLIP [11]61.3579.3585.7555.75 TABLE 10: Results of adopting different retrieval baselines in the adaptive confidence-weighted learning framework on CUHK-PEDES. thesized descriptions and the effectiveness of our proposed adaptive confidence-weighted retrieval learning framework. 4.6.2 Different baseline retrieval models We evaluate the compatibility of the proposed adaptive confidence-weighted learning framework by integrating various baseline retrieval models. As shown in Table 10, the framework demonstrates strong adaptability and robust- ness, consistently achieving competitive performance across different retrieval models. 4.7 Visualization Analysis Figure 9 presents diverse examples illustrating the top-5 retrieval results of both the BLIP baseline and our GTR+, demonstrating effectiveness and superiority of GTR+. Fig- ure 10 visualizes the activation maps for positive image-text pairs using BLIP [11], GTR [12] and our GTR+. Despite the unsupervised setting, GTR+ places greater emphasis on the described object in the image and the key attributes in the text during computation, compared to other methods. 5 CONCLUSION In this paper, we address the unsupervised TBPS task, aiming to remove reliance on expensive human-annotated textual descriptions. To this end, we propose GTR+, a two- stage framework that first generates pseudo-descriptions to compensate for the absence of annotations, and subse- quently trains the retrieval model in a supervised fashion. In the generation stage, we introduce a tiered description IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE14 amanwithblackhairis wearingglasses,awhite hoodedsweatshirt,grey pants,whitesneakersand ablackbag. amaninhisthirtieswith medium-lengthstraightblack hairiswearingablack thermaljacketheisalso wearingapairofblackpants andwhitesneakersheis carryinganorangebackpack. ayoungwomaniswearinga blueandgreyinsulatedcoat withahoodthehoodhasa greyfurliningonitsheis wearingblackleggingsand bluerunningshoeswith whitesolessheisholdinga blacksoftsideluggagebag. Thewomanhasdark mediumlengthhairthat isinaponytailsheis wearalongwhiteshirt withabrownjacketand bluecutoffpantssheis alsowearingwhiteblack tennisshoes themanwearingthedark plaidcoatwiththe hood,apairofbluejeans andbrownshoesis puttingthehandsinthe pocketsandheiswearing apieceofredclothing underthecoat. thewomaniswearinga blackdownjacketanda blackskirtwithsome patternsonitshehas brownhair Baseline Ours Baseline Ours Baseline Ours CUHK-PEDES (Rank@1 →Rank@5) ICFG-PEDES (Rank@1 →Rank@5) Baseline Ours RSTPReid (Rank@1 →Rank@5) QueryQueryQuery Baseline Ours Baseline Ours Fig. 9: Visualization of retrieval results (R@1 to R@5). Each row represents a query-response group. Green and red bounding boxes indicate correct and incorrect matches, respectively. BLIP [11] is used as the baseline. OursBaselineImageGTR GTRBaselineImageOurs Thiswomanhas onabluedressthat goestoherknees, blackandwhite tennisshoesand totingaredbag Query Themanis wearingastripped shortsleevedshirt andhasondark coloredjeans,he alsoiscarryinga shoulderbag Query Fig. 10: Visualization of activation maps. Warmer colors indicate stronger activation during model computation. BLIP [11] is used as the baseline. generation framework that produces fine-grained and stylis- tically diverse captions through a three-tiered sequential process. In the retrieval stage, we propose an adaptive confidence-weighted learning framework that dynamically adjusts each sample’s training contribution based on its confidence score, effectively reducing the negative impact of noisy pseudo texts. Moreover, we contribute LargeFine- Person, a new large-scale dataset featuring high-quality, fine-grained, and diverse textual annotations, which serves as a more realistic and challenging benchmark for TBPS. Extensive experiments demonstrate that GTR+ achieves im- pressive performance on multiple TBPS benchmarks under the unsupervised setting, and also validate the effectiveness of the proposed dataset. REFERENCES [1]Z. Li, L. Jianbo, Y. Shi, J. Chen, S. Huang, L. Tu, F. Shen, and H. Ling, “Exploring the potential of large vision-language models for unsupervised text-based person retrieval,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 5, 2025, p. 5119–5127. [2]Y. Bai, M. Cao, D. Gao, Z. Cao, C. Chen, Z. Fan, L. Nie, and M. Zhang, “Rasa: Relation and sensitivity aware repre- sentation learning for text-based person search,” arXiv preprint arXiv:2305.13653, 2023. [3]D. Jiang and M. Ye, “Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval,” in IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. [4]C. Zhao, X. Lv, S. Dou, S. Zhang, J. Wu, and L. Wang, “Incre- mental generative occlusion adversarial suppression network for person reid,” IEEE Transactions on Image Processing, vol. 30, p. 4212–4224, 2021. [5]M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. Hoi, “Deep learning for person re-identification: A survey and out- look,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 6, p. 2872–2893, 2021. [6]M. Cao, S. Li, J. Li, L. Nie, and M. Zhang, “Image-text retrieval: A survey on recent research and development,” arXiv preprint arXiv:2203.14713, 2022. [7]L. Qu, M. Liu, J. Wu, Z. Gao, and L. Nie, “Dynamic modality interaction modeling for image-text retrieval,” in Proceedings of the 44th International ACM SIGIR Conference on Research and Devel- opment in Information Retrieval, 2021, p. 1104–1113. [8]Y. Qin, Y. Chen, D. Peng, X. Peng, J. T. Zhou, and P. Hu, “Noisy-correspondence learning for text-to-image person re- identification,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024, p. 27 197–27 206. [9]S. Li, M. Cao, and M. Zhang, “Learning semantic-aligned feature representation for text-based person search,” in ICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, p. 2724–2728. [10]Y. Bai, Y. Ji, M. Cao, J. Wang, and M. Ye, “Chat-based person retrieval via dialogue-refined cross-modal alignment,” in Proceed- ings of the Computer Vision and Pattern Recognition Conference, 2025, p. 3952–3962. [11]J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language- image pre-training for unified vision-language understanding and generation,” in International Conference on Machine Learning. PMLR, 2022, p. 12 888–12 900. [12]Y. Bai, J. Wang, M. Cao, C. Chen, Z. Cao, L. Nie, and M. Zhang, “Text-based person search without parallel image-text data,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, p. 757–767. [13]W. Tan, C. Ding, J. Jiang, F. Wang, Y. Zhan, and D. Tao, “Har- nessing the power of mllms for transferable text-to-image person reid,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, p. 17 127–17 137. [14]J. Zuo, J. Hong, F. Zhang, C. Yu, H. Zhou, C. Gao, N. Sang, and J. Wang, “Plip: Language-image pre-training for person rep- resentation learning,” Advances in Neural Information Processing Systems, vol. 37, p. 45 666–45 702, 2024. [15]J. Jiang, C. Ding, W. Tan, J. Wang, J. Tao, and X. Xu, “Modeling thousands of human annotators for generalizable text-to-image IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE15 person re-identification,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, p. 9220–9230. [16]S. Li, C. He, X. Xu, F. Shen, Y. Yang, and H. T. Shen, “Adaptive uncertainty-based learning for text-based person retrieval,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 4, 2024, p. 3172–3180. [17]D. Fu, D. Chen, J. Bao, H. Yang, L. Yuan, L. Zhang, H. Li, and D. Chen, “Unsupervised pre-training for person re- identification,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, 2021, p. 14 750–14 759. [18]G. Song, B. Leng, Y. Liu, C. Hetang, and S. Cai, “Region- based quality estimation network for large-scale person re- identification,” in Proceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018. [19]Z. Shao, X. Zhang, C. Ding, J. Wang, and J. Wang, “Uni- fied pre-training with pseudo texts for text-to-image person re- identification,” in Proceedings of the IEEE/CVF International Con- ference on Computer Vision, 2023, p. 11 174–11 184. [20]S. Yang, Y. Zhou, Z. Zheng, Y. Wang, L. Zhu, and Y. Wu, “Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, p. 4492–4501. [21]D. Fu, D. Chen, H. Yang, J. Bao, L. Yuan, L. Zhang, H. Li, F. Wen, and D. Chen, “Large-scale pre-training for person re- identification with noisy labels,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, p. 2476–2486. [22]S. Li, T. Xiao, H. Li, B. Zhou, D. Yue, and X. Wang, “Person search with natural language description,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, p. 1970–1979. [23]Z. Zheng, L. Zheng, M. Garrett, Y. Yang, M. Xu, and Y.-D. Shen, “Dual-path convolutional image-text embeddings with instance loss,” ACM Transactions on Multimedia Computing, Communica- tions, and Applications (TOMM), vol. 16, no. 2, p. 1–23, 2020. [24]A. Farooq, M. Awais, F. Yan, J. Kittler, A. Akbari, and S. S. Khalid, “A convolutional baseline for person re-identification using vi- sion and language descriptions,” arXiv preprint arXiv:2003.00808, 2020. [25]Y. Chen, G. Zhang, Y. Lu, Z. Wang, and Y. Zheng, “Tipcb: A simple but effective part-based convolutional baseline for text- based person search,” Neurocomputing, vol. 494, p. 171–181, 2022. [26]W. Suo, M. Sun, K. Niu, Y. Gao, P. Wang, Y. Zhang, and Q. Wu, “A simple and robust correlation filtering method for text-based person search,” in Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXV. Springer, 2022, p. 726–742. [27]Y. Wu, H. Wang, M. Wu, M. Cao, and M. Zhang, “Laip: learning local alignment from image-phrase modeling for text-based per- son search,” in 2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2024, p. 1–10. [28]C. Qi, X. Yang, N. Wang, and X. Gao, “Granularity-aware hyper- bolic representation for text-based person search,” IEEE Transac- tions on Information Forensics and Security, 2025. [29]E. Loper and S. Bird, “Nltk: The natural language toolkit,” arXiv preprint cs/0205028, 2002. [30]Z. Wang, Z. Fang, J. Wang, and Y. Yang, “Vitaa: Visual-textual attributes alignment in person search by natural language,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XII 16. Springer, 2020, p. 402–420. [31]Y. Jing, C. Si, J. Wang, W. Wang, L. Wang, and T. Tan, “Pose- guided multi-granularity attention network for text-based person search,” in Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 34, no. 07, 2020, p. 11 189–11 196. [32]K. Zheng, W. Liu, J. Liu, Z.-J. Zha, and T. Mei, “Hierarchical gumbel attention network for text-based person search,” in Pro- ceedings of the 28th ACM International Conference on Multimedia, 2020, p. 3441–3449. [33]C. Gao, G. Cai, X. Jiang, F. Zheng, J. Zhang, Y. Gong, P. Peng, X. Guo, and X. Sun, “Contextual non-local alignment over full- scale representation for text-based person search,” arXiv preprint arXiv:2101.03036, 2021. [34]C. Gao, G. Cai, X. Jiang, F. Zheng, J. Zhang, Y. Gong, F. Lin, X. Sun, and X. Bai, “Conditional feature learning based transformer for text-based person search,” IEEE Transactions on Image Processing, vol. 31, p. 6097–6108, 2022. [35]Z. Ji, J. Hu, D. Liu, L. Y. Wu, and Y. Zhao, “Asymmetric cross- scale alignment for text-based person search,” IEEE Transactions on Multimedia, 2022. [36]Z. Shao, X. Zhang, M. Fang, Z. Lin, J. Wang, and C. Ding, “Learn- ing granularity-unified representations for text-to-image person re-identification,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, p. 5566–5574. [37]S. Yan, J. Liu, N. Dong, L. Zhang, and J. Tang, “Prototypical prompting for text-to-image person re-identification,” in Proceed- ings of the 32nd ACM International Conference on Multimedia, 2024, p. 2331–2340. [38]Z. Lu, R. Lin, Y.-P. Tan, and H. Hu, “Prompt-guided transformer and mllm interactive learning for text-based pedestrian search,” IEEE Transactions on Information Forensics and Security, 2025. [39]J. Park, D. Kim, B. Jeong, and S. Kwak, “Plot: Text-based person search with part slot attention for corresponding part discovery,” in European Conference on Computer Vision.Springer, 2024, p. 474–490. [40]Y. Wu, Z. Yan, X. Han, G. Li, C. Zou, and S. Cui, “Lapscore: language-guided person search via color reasoning,” in Proceed- ings of the IEEE/CVF International Conference on Computer Vision, 2021, p. 1624–1633. [41]Z. Wang, A. Zhu, J. Xue, X. Wan, C. Liu, T. Wang, and Y. Li, “Caibc: Capturing all-round information beyond color for text- based person retrieval,” in Proceedings of the 30th ACM Interna- tional Conference on Multimedia, 2022, p. 5314–5322. [42]X. Han, S. He, L. Zhang, and T. Xiang, “Text-based person search with limited data,” arXiv preprint arXiv:2110.10807, 2021. [43]S. Yan, N. Dong, L. Zhang, and J. Tang, “Clip-driven fine- grained text-image person re-identification,” arXiv preprint arXiv:2210.10276, 2022. [44]D. Wei, S. Zhang, T. Yang, and J. Liu, “Calibrating cross- modal feature for text-based person searching,” arXiv preprint arXiv:2304.02278, 2023. [45]M. Cao, Y. Bai, Z. Zeng, M. Ye, and M. Zhang, “An empirical study of clip for text-based person search,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 1, 2024, p. 465–473. [46]A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning.PMLR, 2021, p. 8748–8763. [47]Y. Qin, C. Chen, Z. Fu, D. Peng, X. Peng, and P. Hu, “Human- centered interactive learning via mllms for text-to-image person re-identification,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, p. 14 390–14 399. [48]Y. Xu, J. Wu, Z. Chen, D. Gao, Y. Yang, Z. Lei, and M. Cao, “Sa-person: Text-based person retrieval with scene-aware re- ranking,” arXiv preprint arXiv:2505.24466, 2025. [49]S. Zhao, C. Gao, Y. Shao, W.-S. Zheng, and N. Sang, “Weakly supervised text-based person re-identification,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, p. 11 395–11 404. [50]Y. Zheng, X. Zhao, C. Lan, X. Zhang, B. Huang, J. Yang, and D. Yu, “Cpcl: Cross-modal prototypical contrastive learning for weakly supervised text-based person re-identification,” arXiv preprint arXiv:2401.10011, 2024. [51]A. Fu, J. Zhao, Y. Zhou, W. Du, R. Yao, and A. El Saddik, “Simi- larity regulation and calibration alignment for weakly supervised text-based person re-identification,” ACM Transactions on Multi- media Computing, Communications and Applications, vol. 21, no. 3, p. 1–19, 2025. [52]T. Gong, J. Wang, and L. Zhang, “Enhancing cross-modal com- pletion and alignment for unsupervised incomplete text-to-image person retrieval,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024, p. 794–802. [53]Y. Zhang, Y. Shang, and H. Li, “Dual-granularity cross-modal identity association for weakly-supervised text-to-person image matching,” arXiv preprint arXiv:2507.06744, 2025. [54]D. Gao, Y. Bai, M. Cao, H. Dou, M. Ye, and M. Zhang, “Semi-supervised text-based person search,” arXiv preprint arXiv:2404.18106, 2024. [55]Y. Jing, W. Wang, L. Wang, and T. Tan, “Cross-modal cross- domain moment alignment network for person search,” in Pro- IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE16 ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, p. 10 678–10 686. [56]Z. Li, J. Li, Y. Shi, H. Ling, J. Chen, R. Wang, and S. Huang, “Cross-modal generation and alignment via attribute-guided prompt for unsupervised text-based person retrieval,” in Proceed- ings of the International Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intelligence Organization, 2024, p. 1047–1055. [57]L. H. Li, H. You, Z. Wang, A. Zareian, S.-F. Chang, and K.-W. Chang, “Unsupervised vision-and-language pre-training without parallel images and captions,” in Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021, p. 5339–5350. [58]M. Zhou, L. Yu, A. Singh, M. Wang, Z. Yu, and N. Zhang, “Unsupervised vision-and-language pre-training via retrieval- based multi-granular alignment,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, p. 16 485–16 494. [59]C. Chen, P. Li, M. Sun, and Y. Liu, “End-to-end unsuper- vised vision-and-language pre-training with referring expression matching,” in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022, p. 10 799–10 810. [60]T. Wang, W. Jiang, Z. Lu, F. Zheng, R. Cheng, C. Yin, and P. Luo, “Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix,” in International Conference on Machine Learn- ing. PMLR, 2022, p. 22 680–22 690. [61]Y. Dong, Y. Zhang, L. Ma, Z. Wang, and J. Luo, “Unsupervised text-to-image synthesis,” Pattern Recognition, vol. 110, p. 107573, 2021. [62]Y. Feng, L. Ma, W. Liu, and J. Luo, “Unsupervised image caption- ing,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, p. 4125–4134. [63]J. Gu, S. Joty, J. Cai, H. Zhao, X. Yang, and G. Wang, “Unpaired image captioning via scene graph alignments,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, p. 10 323–10 332. [64]D. Guo, Y. Wang, P. Song, and M. Wang, “Recurrent relational memory network for unsupervised image captioning,” in Pro- ceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, 2021, p. 920–926. [65]H. Ben, Y. Pan, Y. Li, T. Yao, R. Hong, M. Wang, and T. Mei, “Unpaired image captioning with semantic-constrained self- learning,” IEEE Transactions on Multimedia, vol. 24, p. 904–916, 2021. [66]Y. Huang, Y. Wang, Y. Zeng, J. Huang, Z. Chai, and L. Wang, “Un- paired image-text matching via multimodal aligned conceptual knowledge,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. [67]P. Zhu, X. Wang, L. Zhu, Z. Sun, W.-S. Zheng, Y. Wang, and C. Chen, “Prompt-based learning for unpaired image caption- ing,” IEEE Transactions on Multimedia, vol. 26, p. 379–393, 2023. [68]Y. Huang, Y. Wang, Y. Zeng, and L. Wang, “Mack: Multimodal aligned conceptual knowledge for unpaired image-text match- ing,” Advances in Neural Information Processing Systems, vol. 35, p. 7892–7904, 2022. [69]Y. Qi, W. Zhao, and X. Wu, “Relational distant supervision for image captioning without image-text pairs,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 5, 2024, p. 4524–4532. [70]Z. Zeng, Y. Xie, H. Zhang, C. Chen, B. Chen, and Z. Wang, “Meacap: Memory-augmented zero-shot image captioning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, p. 14 100–14 110. [71]W. Li, L. Zhu, L. Wen, and Y. Yang, “Decap: Decoding clip latents for zero-shot captioning via text-only training,” arXiv preprint arXiv:2303.03032, 2023. [72]Z. Liu, J. Liu, X. Ling, Q. Huang, and J. Wang, “Synthesize then align: Modality alignment augmentation for zero-shot image captioning with synthetic data,” Knowledge-Based Systems, vol. 315, p. 113274, 2025. [73]J. Luo, J. Chen, Y. Li, Y. Pan, J. Feng, H. Chao, and T. Yao, “Unleashing text-to-image diffusion prior for zero-shot image captioning,” in European Conference on Computer Vision. Springer, 2024, p. 237–254. [74]Z. Liu, J. Liu, and F. Ma, “Improving cross-modal alignment with synthetic pairs for text-only image captioning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 4, 2024, p. 3864–3872. [75]L. Qiu, S. Ning, and X. He, “Mining fine-grained image-text alignment for zero-shot captioning via text-only training,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 5, 2024, p. 4605–4613. [76]J. Fei, T. Wang, J. Zhang, Z. He, C. Wang, and F. Zheng, “Transferable decoding with visual entities for zero-shot image captioning,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, p. 3136–3146. [77]X. Yang, H. Zhang, and J. Cai, “Deconfounded image captioning: A causal retrospect,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 11, p. 12 996–13 010, 2021. [78]B. Z. Yao, X. Yang, L. Lin, M. W. Lee, and S.-C. Zhu, “I2t: Image parsing to text description,” Proceedings of the IEEE, vol. 98, no. 8, p. 1485–1508, 2010. [79]X. Yang, H. Zhang, and J. Cai, “Auto-encoding and distilling scene graphs for image captioning,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 5, p. 2313–2327, 2020. [80]O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A neural image caption generator,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, p. 3156–3164. [81]M. Cornia, L. Baraldi, and R. Cucchiara, “Explaining transformer- based image captioning models: An empirical analysis,” AI Com- munications, vol. 35, no. 2, p. 111–129, 2022. [82]I. Al Badarneh, B. H. Hammo, and O. Al-Kadi, “An ensemble model with attention based mechanism for image captioning,” Computers and Electrical Engineering, vol. 123, p. 110077, 2025. [83]M. Barraco, S. Sarto, M. Cornia, L. Baraldi, and R. Cucchiara, “With a little help from your own past: prototypical memory networks for image captioning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, p. 3021–3031. [84]J. Hessel, A. Holtzman, M. Forbes, R. L. Bras, and Y. Choi, “Clip- score: A reference-free evaluation metric for image captioning,” arXiv preprint arXiv:2104.08718, 2021. [85]H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi ` ere, N. Goyal, E. Hambro, F. Azhar et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023. [86]L. Chen, J. Li, X. Dong, P. Zhang, C. He, J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” in European Conference on Computer Vision. Springer, 2024, p. 370–387. [87]J. Chen, Z. Xu, X. Pan, Y. Hu, C. Qin, T. Goldstein, L. Huang, T. Zhou, S. Xie, S. Savarese et al., “Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset,” arXiv preprint arXiv:2505.09568, 2025. [88]X. Yang, Y. Wu, M. Yang, H. Chen, and X. Geng, “Exploring diverse in-context configurations for image captioning,” Advances in Neural Information Processing Systems, vol. 36, p. 40 924–40 943, 2023. [89]D. Bucciarelli, N. Moratelli, M. Cornia, L. Baraldi, and R. Cuc- chiara, “Personalizing multimodal large language models for im- age captioning: an experimental analysis,” in European Conference on Computer Vision. Springer, 2025, p. 351–368. [90]H. Dong, J. Li, B. Wu, J. Wang, Y. Zhang, and H. Guo, “Bench- marking and improving detail image caption,” arXiv preprint arXiv:2405.19092, 2024. [91]Z. Huang, G. Niu, X. Liu, W. Ding, X. Xiao, H. Wu, and X. Peng, “Learning with noisy correspondence for cross-modal match- ing,” Advances in Neural Information Processing Systems, vol. 34, p. 29 406–29 419, 2021. [92]H. Han, K. Miao, Q. Zheng, and M. Luo, “Noisy correspondence learning with meta similarity correction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, p. 7517–7526. [93]Z. Dang, M. Luo, J. Wang, C. Jia, H. Han, H. Wan, G. Dai, X. Chang, and J. Wang, “Disentangled noisy correspondence learning,” IEEE Transactions on Image Processing, 2025. [94]H. Shi, M. Liu, X. Mu, X. Song, Y. Hu, and L. Nie, “Breaking through the noisy correspondence: A robust model for image- text matching,” ACM Transactions on Information Systems, vol. 42, no. 6, p. 1–26, 2024. [95]M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al., IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE17 “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023. [96]P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge et al., “Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,” arXiv preprint arXiv:2409.12191, 2024. [97]Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liu et al., “Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling,” arXiv preprint arXiv:2412.05271, 2024. [98]K. Niu, Y. Huang, W. Ouyang, and L. Wang, “Improving description-based person re-identification by multi-granularity image-text alignments,” IEEE Transactions on Image Processing, vol. 29, p. 5542–5556, 2020. [99]D. Jiang and M. Ye, “Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, p. 2787–2797. [100] M. Sun, W. Suo, P. Wang, K. Niu, L. Liu, G. Lin, Y. Zhang, and Q. Wu, “An adaptive correlation filtering method for text-based person search,” International Journal of Computer Vision, vol. 132, no. 10, p. 4440–4455, 2024. [101] Z. Ding, C. Ding, Z. Shao, and D. Tao, “Semantically self-aligned network for text-to-image part-aware person re-identification,” arXiv preprint arXiv:2107.12666, 2021. [102] A. Zhu, Z. Wang, Y. Li, X. Wan, J. Jin, T. Wang, F. Hu, and G. Hua, “Dssl: Deep surroundings-person separation learning for text-based person retrieval,” in Proceedings of the 29th ACM International Conference on Multimedia, 2021, p. 209–217. [103] S. Yang, Y. Wang, Y. Li, L. Zhu, and Z. Zheng, “Minimizing the pretraining gap: Domain-aligned text-based person retrieval,” arXiv preprint arXiv:2507.10195, 2025. Mang Ye (Senior Member, IEEE) received the PhD degree in computer science from Hong Kong Baptist University, in 2019. He is currently a full professor with the School of Computer Science, Wuhan University, Wuhan, China. He has published more than 100 articles in top-tier venues. He serves as the associate editor for IEEE Transactions on Image Processing, IEEE Transactions on Information Forensics and Se- curity, the Journal of Electronic Imaging, and CAAI Transactions on Intelligence Technology. His research interests focus on computer vision, pattern recognition, and federated learning. Yucheng Ji received the B.E. degree from the School of Future Science and Engineering, Soo- chow University, Suzhou, China. He is currently working toward the ME degree with the School of Computer Science and Technology, Soochow University. His research interests include cross- modal retrieval and person re-identification. Yang Bai received the M.E. degree from the School of Computer Science and Technology, Soochow University, Suzhou, China. He is cur- rently pursuing the Ph.D. degree with the School of Computer Science, Wuhan University. His re- search interests include cross-modal retrieval and person re-identification. Min Cao received her Ph.D. degree in pattern recognition and intelligent systems from the In- stitute of Automation, Chinese Academy of Sci- ences, Beijing, China, in 2020. In March 2020, she became a member of the computer science and technology school at Soochow University, where she is currently an Associate Professor. She was a visiting scholar in computer graphics research, Fraunhofer-Gesellschaft, Darmstadt, Germany, in 2018. Her research interests in- clude cross-modal vision-language learning and person re-identification. Siyuan Chai received his M.E. degree from the School of Computer Science and Technol- ogy, Jilin University, Jilin, China. Since 2023, he has been serving as the General Manager of the Business Technology Center at Zhipu AI. His research interests focus on the applications of large language models, including Retrieval- Augmented Generation (RAG), Agents, Natural Language to SQL (NL2SQL) conversion, and multimodal large language models. Bo Du (Senior Member, IEEE) received the PhD degree in photogrammetry and remote sensing from the State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, China, in 2010. He is a professor with the School of Com- puter Science, Wuhan University. He has more than 60 research articles published in the IEEE TGRS, TIP, JSTARS, and GRSL. Thirteen of them are ESI hot articles or highly cited arti- cles. His major research interests include pattern recognition, hyperspectral image processing, machine learning, and signal processing. He was a recipient of the Distinguished Paper Award from IJCAI 2018, the Best Paper Award of the IEEE Whispers 2018, the Champion Award of the IEEE Data Fusion Contest 2018, the Best Reviewer Award from the IEEE GRSS for his service to the IEEE Journal of Selected Topics in Earth Observations and Applied Remote Sensing, in 2011, and the ACM rising star awards for his academic progress, in 2015. He was the session chair of the IGARSS 2018/2016 and the 4th IEEE GRSS Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing. Min Zhang (ACL Fellow, and AAIA Fellow) re- ceived the bachelor’s and PhD degrees from the Harbin Institute of Technology, in 1991 and 1997, respectively. He is a distinguished profes- sor with Soochow University (China). His current research interests include machine translation, natural language processing, large model and AI.