Paper deep dive
Visual-Prompt Guided Wildlife Instance-Level Recognition
Mufhumudzi Muthivhi, Jiahao Huo, Terence van Zyl, Fredrik Gustafsson
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 8/20/2026, 4:10:24 AM
Summary
This paper proposes a one-stage end-to-end visual-prompt guided framework for wildlife instance-level recognition, combining DINOv2 for spatial geometry and MegaDescriptor for fine-grained identity features. The model uses cross-attention mechanisms to query scene latent spaces with identity prompts, achieving a competitive mean average precision (mAP) of 30.584% compared to a two-stage baseline of 44.89%.
Entities (8)
Relation Signals (8)
Visual-Prompt Guided Wildlife Instance-Level Recognition → proposes → one-stage end-to-end detection and re-identification model
confidence 98% · We propose a one-stage end-to-end detection and re-identification model that performs identity searching within the latent space.
Visual-Prompt Guided Wildlife Instance-Level Recognition → uses → DinoV2
confidence 98% · We adopt DINOv2 for robust spatial geometry... We use DINOv2 [6] to process the full scene image.
Visual-Prompt Guided Wildlife Instance-Level Recognition → uses → MegaDescriptor
confidence 98% · We adopt... MegaDescriptor for wildlife re-identification.
Visual-Prompt Guided Wildlife Instance-Level Recognition → achievesmap → 30.584%
confidence 95% · Preliminary findings reflect a competitive mean average precision score of 30.584%...
MegaDetectorV6 + MegaDescriptor → achievesmap → 44.89%
confidence 95% · ...compared to the state-of-the-art two stage approach of 44.89%.
Visual-Prompt Guided Wildlife Instance-Level Recognition → comparesagainst → MegaDetectorV6 + MegaDescriptor
confidence 95% · The results shown is evaluated on the combined test set... compared it to the two-stage approach that uses MegadetectorV6... with MegaDescriptor
Visual-Prompt Guided Wildlife Instance-Level Recognition → evaluateson → ATRW
confidence 95% · We specifically utilize the GiraffeZebraID [7] and ATRW [4] datasets...
Visual-Prompt Guided Wildlife Instance-Level Recognition → evaluateson → GiraffeZebraID
confidence 95% · We specifically utilize the GiraffeZebraID [7] and ATRW [4] datasets...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Fine-grained wildlife re-identification remains a challenging area in research. Current state-of-the-art approaches apply a detection and re-identification pipeline. We propose a one-stage end-to-end detection and re-identification model that performs identity searching within the latent space. We adopt DINOv2 for robust spatial geometry and MegaDescriptor for wildlife re-identification. We enhance latent queries with prompt re-identification features. A detection decoder queries the scene latent space to establish object boundaries around the target identity. Preliminary findings reflect a competitive mean average precision score of 30.584% compared to the state-of-the-art two stage approach of 44.89%. Qualitative results depict effective bounding and identification of animal identities.
Tags
Links
- Source: https://arxiv.org/abs/2608.18246v1
- Canonical: https://arxiv.org/abs/2608.18246v1
Trouble viewing inline? Open PDF directly →
Full Text
10,826 characters extracted from source content.
Expand or collapse full text
Visual-Prompt Guided Wildlife Instance-Level Recognition Mufhumudzi Muthivhi 1 , Jiahao Huo 1 , Terence van Zyl 1 , and Fredrik Gustafsson 2 1 University of Johannesburg, South Africa mmuthivhi,216045414,tvanzyl@uj.ac.za 2 Linköping University, Sweden fredrik.gustafsson@liu.se Abstract. Fine-grained wildlife re-identification remains a challenging area in research. Current state-of-the-art approaches apply a detection and re-identification pipeline. We propose a one-stage end-to-end detec- tion and re-identification model that performs identity searching within the latent space. We adopt DINOv2 for robust spatial geometry and MegaDescriptor for wildlife re-identification. We enhance latent queries with prompt re-identification features. A detection decoder queries the scene latent space to establish object boundaries around the target iden- tity. Preliminary findings reflect a competitive mean average precision score of 30.584% compared to the state-of-the-art two stage approach of 44.89%. Qualitative results depict effective bounding and identification of animal identities. Keywords: instance recognition· re-identification· grounding· wildlife 1 Introduction Generic object detection models categorize wildlife at the species level. Con- versely, instance-level wildlife re-identification requires the discrimination of fine- grained visual patterns. The model is tasked with learning from the unique stripe configurations of a zebra or the spot patterns of a cheetah. Current state-of-the- art approaches deploy an isolated detection and re-identification pipeline using two separate models [8]. An object detection model isolates individuals in a scene, crops the bounding box pixels, resizes them to a fixed resolution, and passes them to an independent re-identification feature extractor for similar- ity matching against a database [2]. Recent advancements in promptable vision foundation models and visual grounding present a compelling alternative [5]. We could instead view the database as a prompt to query the scene directly to locate a target. However, adapting single-stage visual grounding architec- tures for instance-level re-identification introduces a training mismatch. A single transformer decoder layer is forced to simultaneously learn the global scenery to extract geometric boundaries of an object and the local texture patterns of a spe- cific identity. To resolve this, we propose an end-to-end visual prompting task. arXiv:2608.18246v1 [cs.CV] 18 Aug 2026 2Muthivhi et al. A frontier foundation model extracts robust spatial features and a specialist re- identification model extracts fine-grained features. Preliminary findings depict a competitive mean average precision score (mAP) over a two-stage approach. Qualitative results also depict tight bounding boxes around target identities in dense herd, occluded environments and distant subjects. DINOv2 MegaDescriptor Multi-Scale Scene Encoder Queries Cross-Attention Detector FFN Feature Enhancer Prompt Self-Attention Query-to-Prompt Cross-Attention Query Self-Attention Prompt-to-Query Cross-Attention ❄ ❄ Self-Attention Deformable Self-Attention scene to prompt cross attention prompt to scene cross attention bbox bbox bbox bbox . . Cross Attention Detector Trainable ❄ Frozen Embeddings Fig. 1: Overview of the proposed framework. DINOv2 encodes the full scene image to capture spatial features and MegaDescriptor encodes reference crops into identity features used as visual prompts for target matching. 2 Methodology The proposed framework in Figure 1 consists of three primary components. First, we extract scene and re-identification features from a frozen encoder. Next, we enhance the object queries with prompt re-identification features to enrich our region proposals. Lastly, we use a cross-attention detector to search for the target identity. Multi-Scale Scene Encoder We use DINOv2 [6] to process the full scene image. To handle extreme scale variations in imagery, we adopt a similar strategy used by Grounding DINO [5]. We extract intermediate feature maps from layers 3, 6 and 10 within the transformer. These features are projected into a feature pyramid using varying convolutional strides, spatially interpolated to a uniform resolution, concatenated, and subsequently compressed via a 1× 1 convolution. Visual prompt The visual prompt is obtained through a pretrained fine-grained wildlife re-identification model, MegaDescriptor [2]. Feature Enhancer We initialize a set of learnable object queries as the region proposals. The feature enhancement layer ensures that the queries cross-attend directly with the visual prompt features. We use both self and cross attention to Visual-Prompt Guided Wildlife Instance-Level Recognition3 enrich our region proposals (queries) with identity-related information from our visual prompts. Cross-Attention Decoder The enriched region proposals attend to the multi- scale scene features via Deformable Attention [9]. The decoder outputs a set of N cross-attended queries. The cross-attended queries are passed through a feed-forward network to encapsulate the detected animal with a bounding box. An additional objectiveness score is also retrieved, indicating the presence of a valid foreground animal. Loss Function This work proposes the optimization of four objectives, such that: L = λ obj L FocalLoss + λ L1 L L1 + λ GIoU L GIoU + λ div L div (1) is the detection loss. L F L is the Sigmoid Focal Loss over the objectness score obtained from the N queries. L L1 and L GIoU is the L1 loss and Generalized Intersection over Union of the predicted and ground truth bounding boxes. L div encourages diversity of the embeddings of the queries produced by the feature enhancement layer. Fig. 2: Qualitative visual grounding results of the proposed architecture. For each sce- nario, the target identity prompts are displayed on the left, and the scene interrogation results are on the right. 3 Experimental Setup Datasets We use a curated re-identification dataset from WildlifeDatasets [2]. We specifically utilize the GiraffeZebraID [7] and ATRW [4] datasets because they come with the full-scene image and the bounding box metadata. Implmentation We employ a ViT-S/14 variant of DINOv2 for the multi-scale scene encoder and a MegaDescriptor-T variant for the prompt encoder. Scene images are resized to 518× 518 to align with the DINOv2 patch configuration, while prompt crops are resized to 224× 224. We selectively unfreeze the final transformer block and normalization layers of both encoders during training, al- lowing their distinct latent manifolds to synchronize. We use PyTorch Lightning for efficient training [3]. 4Muthivhi et al. Hyperparameters: The network is optimized end-to-end using the AdamW op- timizer with a base learning rate of 1× 10 −4 and a weight decay of 1× 10 −4 . We employ a MultiStepLR learning rate scheduler, which decays the rate by a factor of γ = 0.1. The objectness, identity matching, bounding box L1 distance, and Generalized IoU is set to 2.0, 2.0, 5.0, and 2.0, respectively. 4 Preliminary Results mAP@[0.5:0.95]AP50AP60AP70AP90 0 10 20 30 40 50 60 70 80 Percentage (%) Score 44.89% 80.23% 73.55% 57.75% 3.85% 30.84% 71.28% 57.52% 33.74% 0.38% Two-stage vs One-stage Two-Stage (MegaDetectorV6 + MegaDescriptor) Ours (End-to-End Latent Grounding) Fig. 3: Our End-to-End Latent Grounding method (one stage) compared to two stage method using MegadetectorV6 [1] with MegaDescriptor [2]. The results shown is eval- uated on the combined test set from GiraffeZebraID [7] and ATRW [4] We provide qualitative preliminary findings in Figure 2. The proposed frame- work produces bounding boxes that encapsulate the target animal. Our model can handle herds, domain shifts, occlusion, distance, and camouflage. We present results on retrieval performance (ReID) in Figure 3 using our end-to-end la- tent grounding method and compared it to the two-stage approach that uses MegadetectorV6 for detection, crop and followed by MegaDescriptor to embed the cropped images to perform retrieval. The results show the potential of the proposed approach even though it has not yet achieved state-of-the-art perfor- mance. However, the proposed approach is much more efficient. 5 Conclusion This paper proposes a framework that combines visual prompt along with global scene features to perform end-to-end object detection together with re-identification. The preliminary findings suggest that visual identity prompts can reduce the re- liance on traditional two-stage detect and crop followed by ReID pipelines by allowing the model to localize and match a target directly within the full scene image. We believe continued work in this direction could lead to state-of-the-art performance. Visual-Prompt Guided Wildlife Instance-Level Recognition5 References 1. Beery, S., Morris, D., Yang, S.: Efficient pipeline for camera trap image review (2019) 2. Čermák, V., Picek, L., Adam, L., Papafitsoros, K.: Wildlifedatasets: An open-source toolkit for animal re-identification. In: Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision. p. 5953–5963 (2024) 3. Falcon, W., team, T.P.L.: Pytorch lightning. https://doi.org/10.5281/zenodo. 3828935, https://github.com/Lightning-AI/pytorch-lightning 4. Li, S., Li, J., Tang, H., Qian, R., Lin, W.: Atrw: A benchmark for amur tiger re- identification in the wild. arXiv preprint arXiv:1906.05586 (2019) 5. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 6. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) 7. Parham, J., Crall, J., Stewart, C., Berger-Wolf, T., Rubenstein, D.I.: Animal pop- ulation censusing at scale with citizen science and photographic identification. In: AAAI spring symposium-technical report (2017) 8. Wu, Y., Zhao, D., Li, Y., Alajas, M., Glen, A.S., Zhang, J., Dobbie, G., Wilson, D., Koh, Y.S.: Overcoming fine-grained visual challenges in animal re-identification via semantic feature alignment. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. p. 371–381 (2026) 9. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 4794–4803 (2022)