Paper deep dive
Integrating Multimodal Large Language Model Knowledge into Amodal Completion
Heecheol Yun, Eunho Yang
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/31/2026, 2:24:24 AM
Summary
AmodalCG is a novel framework for amodal completion that integrates Multimodal Large Language Models (MLLMs) to provide geometric and semantic guidance for reconstructing occluded objects. It features a selective invocation mechanism to manage computational costs, a multi-scale expansion strategy to handle geometric uncertainty, and uses visual generative models to refine completions, outperforming existing methods in amodal segmentation and object recognition.
Entities (5)
Relation Signals (3)
AmodalCG → uses → Stable Diffusion
confidence 100% · Stable Diffusion v2 inpainting model [17] is used for the Inpainting Module.
AmodalCG → utilizes → MLLM
confidence 100% · AmodalCG, a novel framework that harnesses the real-world knowledge of Multimodal Large Language Models (MLLMs)
MLLM → providesguidancefor → Amodal Completion
confidence 95% · harnesses the real-world knowledge of Multimodal Large Language Models (MLLMs) to guide amodal completion.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:With the widespread adoption of autonomous vehicles and robotics, amodal completion, which reconstructs the occluded parts of people and objects in an image, has become increasingly crucial. Just as humans infer hidden regions based on prior experience and common sense, this task inherently requires physical knowledge about real-world entities. However, existing approaches either depend solely on the image generation ability of visual generative models, which lack such knowledge, or leverage it only during the segmentation stage, preventing it from explicitly guiding the completion process. To address this, we propose AmodalCG, a novel framework that harnesses the real-world knowledge of Multimodal Large Language Models (MLLMs) to guide amodal completion. Our framework first assesses the extent of occlusion to selectively invoke MLLM guidance only when the target object is heavily occluded. If guidance is required, the framework further incorporates MLLMs to reason about both the (1) extent and (2) content of the missing regions. Finally, a visual generative model integrates these guidance and iteratively refines imperfect completions that may arise from inaccurate MLLM guidance. Experimental results on various real-world images show impressive improvements compared to all existing works, suggesting MLLMs as a promising direction for addressing challenging amodal completion.
Tags
Links
- Source: https://arxiv.org/abs/2603.28333v1
- Canonical: https://arxiv.org/abs/2603.28333v1
Trouble viewing inline? Open PDF directly →
Full Text
53,563 characters extracted from source content.
Expand or collapse full text
Integrating Multimodal Large Language Model Knowledge into Amodal Completion Heecheol Yun 1 Eunho Yang 1,2 1 KAIST 2 AITRICS yoon6503, eunhoy@kaist.ac.kr Previous methods (No Guidance) MLLM Query “Front section of a bus with a large windshield, side mirror, and colorful stripes on the side.” 1. Resizing inpainting mask Need guidance? Ours (with MLLM Guidance) 2. Inferring occluded regions Extraneous generation Unnatural output“Chicken” “Bus” Figure 1. Our method selectively leverages MLLMs to reason about the extent and content of occluded parts. Incorporating them into amodal completion effectively enhances performance. Abstract With the widespread adoption of autonomous vehicles and robotics, amodal completion, which reconstructs the oc- cluded parts of people and objects in an image, has become increasingly crucial. Just as humans infer hidden regions based on prior experience and common sense, this task in- herently requires physical knowledge about real-world en- tities. However, existing approaches either depend solely on the image generation ability of visual generative mod- els, which lack such knowledge, or leverage it only during the segmentation stage, preventing it from explicitly guid- ing the completion process. To address this, we propose AmodalCG, a novel framework that harnesses the real- world knowledge of Multimodal Large Language Models (MLLMs) to guide amodal completion. Our framework first assesses the extent of occlusion to selectively invoke MLLM guidance only when the target object is heavily occluded. If guidance is required, the framework further incorporates MLLMs to reason about both the (1) extent and (2) con- tent of the missing regions. Finally, a visual generative model integrates these guidance and iteratively refines im- perfect completions that may arise from inaccurate MLLM guidance. Experimental results on various real-world im- ages show impressive improvements compared to all exist- ing works, suggesting MLLMs as a promising direction for addressing challenging amodal completion. 1. Introduction Imagine a situation where a desired object in a photo is un- intentionally obscured by other foregrounds, preventing us from obtaining its full appearance. Amodal completion [6] is a task designed for such scenarios, with the goal of recon- structing the whole object based on its visible parts. In daily life, occlusion occurs frequently, making amodal comple- tion highly valuable in a range of downstream applications, such as autonomous vehicles and robotics. Similar to how humans infer the hidden parts of objects, amodal completion can greatly benefit from rich common- sense knowledge about real-world entities. However, exist- ing methods [15, 21] overlook this and heavily rely on gen- eration capabilities of visual generative models [17] without providing sufficient guidance about the occluded regions. MC Diffusion [21] proposes a training-free approach that adapts the denoising process of Stable Diffusion (SD) for amodal completion, using only category-level text prompts 1 arXiv:2603.28333v1 [cs.CV] 30 Mar 2026 ImageModal ImageSD Inpaintingpix2gestaltMC DiffusionOurs OWAAC Figure 2. Completion results when meaningful parts of the target object are occluded. Stable Diffusion (SD) inpainting [17] often generates objects other than the target object. Existing amodal completion methods [1, 15, 21] lack an understanding of what should be generated for the missing parts. In contrast, our method provides explicit guidance on what should be reconstructed. as guidance. pix2gestalt [15] fine-tunes SD on synthetic amodal completion datasets, where SD is conditioned on the input image and a modal mask. Although OWAAC [1] recently proposes to employ MLLM [8] in its framework, its role is restricted to segmenting the target object from ab- stract user queries, while SD remains conditioned solely on category-level text prompts. Consequently, these methods often produce unnatural completions in challenging occlu- sion scenarios and require multiple sampling attempts with different seeds to obtain satisfactory results. As illustrated in Figure 2, existing methods generate an object entirely un- related to the bicycle or fail to determine whether a person is sitting or standing based on the image context. To overcome this limitation, we propose AmodalCG (Amodal Completion via MLLM Guidance), a novel framework that effectively leverages the rich physical-world knowledge embedded in Multimodal Large Language Mod- els (MLLMs) [11, 13, 14, 19] for amodal completion. Specifically, AmodalCG identifies and integrates two key types of MLLM-derived guidance about the occluded con- tent. First, the MLLM generates geometric guidance, esti- mating the true extent of the occluded regions. This guid- ance is necessary as SD becomes prone to erroneous com- pletions when the inpainting mask is excessively larger than the actual object. To address this, AmodalCG employs MLLM to estimate the full extent of the target object and uses this prediction to resize the inpainting mask accord- ingly. This offers explicit cues on how much of the object should be reconstructed, preventing over-extended comple- tion and unintended content generation. The second is se- mantic guidance, which provides a detailed textual descrip- tion of what should be generated in the occluded area. Once the inpainting mask is resized to fit the full object, the MLLM infers the appropriate content for the occluded re- gion. This description is then used as a text prompt for SD, giving it explicit guidance on what needs to be filled in. However, incorporating MLLM guidance into amodal completion presents two key challenges. First, generating MLLM guidance for every sample can be inefficient, as some cases—such as those with minimal occlusion—can already be effectively completed without detailed guid- ance. Second, the inherent ambiguity of amodal completion makes it difficult for MLLMs to produce accurate predic- tions about the hidden regions, particularly when estimating the size of the full target object. To address these challenges, we introduce the following two strategies. First, before invoking a large-scale MLLM to generate detailed guidance, a lightweight model is used to assess the degree of occlusion and selectively triggers the large model. For samples with minimal occlusion, the framework proceeds without calling the large model, thereby reducing computational cost. Second, to alleviate the difficulty of estimating the extent of occluded regions, we adopt a multi-scale expansion strategy. Instead of pro- ducing a single estimate, the MLLM predicts multiple can- didate scales for the full target object. Then, starting from 2 the tightest prediction, our framework progressively verifies whether the target object can be fully reconstructed within each predicted region, selecting the most suitable predic- tion for completion. This multi-scale strategy increases the chances of the MLLM predicting an accurate object size, thereby facilitating reliable completion. Our method enables amodal completion for open-world objects without additional training. It is simple yet highly effective, improving both amodal segmentation and oc- cluded object recognition. Our contributions are summa- rized as follows: • We propose AmodalCG, a framework that selectively in- tegrates the rich common-sense knowledge of MLLMs into challenging amodal completion. • We identify two key types of MLLM-derived guidance and alleviates the computational and uncertainty chal- lenges when incorporating these guidance signals through selective large-model invocation and a multi-scale expan- sion strategy. • Our method improves amodal segmentation by 5.49% and occluded object recognition by 2.92% compared to the baselines, highlighting MLLMs as a promising solu- tion for challenging amodal completion. 2. Related work Amodal Completion. Early studies on amodal completion primarily focused on training task-specific models. How- ever, this training inherently requires ground-truth appear- ance for the invisible regions, which are costly to obtain. Consequently, these approaches are typically trained on nar- row domains such as vehicles [10, 22], humans [26, 27], or indoor scenes [3, 4], resulting in limited generalization to categories outside the training data. To overcome the limitations of these datasets, recent studies [1, 15, 21] have leveraged the power of large-scale diffusion models, such as Stable Diffusion (SD) [17], which are trained on massive datasets [18]. These approaches utilize SD to directly complete the appearance of oc- cluded objects. pix2gestalt [15] fine-tunes SD on synthetic datasets curated for amodal completion. MC Diffusion [21] bypasses the expensive fine-tuning stage by proposing a training-free pipeline that first identifies occluders from the segmentation masks [7, 12] and then inpaints the occlud- ers’ regions using SD. To prevent unintended generations, it clusters intermediate features of SD to retain only those fea- tures similar to the target object. Under the similar frame- work, OWAAC [1] enhances the completion performance by employing more accurate category-level text prompts and refining inpainting masks through morphological op- erations. OWAAC also employs MLLM [8] into its frame- work, whose role is restricted to segmenting the target ob- ject and does not influence the completion process. In con- trast, our method incorporates the rich real-world knowl- edge of MLLMs directly into the generation process, pro- viding detailed and explicit guidance to SD for completing occluded regions. 3. AmodalCG: Amodal Completion via MLLM Guidance Given an input image I along with a modal mask for the target object M modal and, in some cases, its semantic cate- gory P cat , recent amodal completion approaches utilize vi- sual generative models, such as Stable Diffusion (SD) [17], to reconstruct the occluded regions. Under the same set- ting, we propose AmodalCG, a framework that leverages the rich real-world knowledge of MLLMs to guide the com- pletion process. Our method consists of five main com- ponents. First, Sec. 3.1 introduces the Occluder Detec- tion Module, which identifies occluders to form the in- painting mask M inpaint . Next, Sec. 3.2 describes the Guid- ance Decision Module, which determines whether to invoke MLLM guidance for reconstructing the target object. We then present two modules that generate MLLM-derived in- formation about the occluded regions: the Geometric Guid- ance Module (Sec. 3.3) and the Semantic Guidance Module (Sec. 3.4). Finally, Sec. 3.5 describes the Inpainting Mod- ule, which integrates both types of MLLM guidance into the completion process through multi-scale expansion. All prompts used in our framework are provided in Appendix Sec. 6. Fig. 3 illustrates an overview of our pipeline. 3.1. Occluder Detection Given an input image I and a modal mask M modal , the Occluder Detection Module outputs the inpainting mask M inpaint , defined as the union of the occluder masks [1, 21]. To identify occluders, we first perform semantic segmenta- tion [7] on I and then use a geometric order prediction net- work [9] to determine the occlusion order of each segment. 3.2. Selective Invocation of MLLM Guidance The Guidance Decision Module selectively invokes MLLM guidance based on the level of occlusion. Since samples with minimal occlusion can be adequately recon- structed without additional guidance, our framework omits generating detailed MLLM guidance when the target object is regarded as minimally occluded. Unlike reasoning about occluded regions, deciding whether an object is nearly com- plete is a simpler task. Therefore, we employ a smaller-scale MLLM for the Guidance Decision Module to assess the ne- cessity of MLLM guidance. Given an image of the isolated target object on a white background, the module outputs two pieces of information in JSON format: (1) a binary indica- tor specifying whether MLLM guidance is required, and (2) the category of the target object P cat . In Secs. 3.3 and 3.4, we describe how this information is subsequently used to generate geometric and semantic guidance. 3 Query object: Doll “A doll with a colorful dress, intricate patterns, and delicate lace, holding a small bouquet of flowers.” Tight: [240, 80, 400, 320] Moderate: [220, 60, 420, 360] Coarse: [200, 40, 440, 400] Describe the occluded part of the <Doll>... Predict the size of the full <Doll>... Tight Moderate Coarse Tight Moderate Guidance Decision Module Need guidance? Larger MLLM Larger MLLM Prompt: Doll Yes No 1st generation 2nd generation Touch tight mask boundary?→ Yes Proceed with next-scale mask Touch moderate mask boundary?→ No Stop generation Inpainting mask: Add a small fixed margin to the modal bounding box Input Figure 3. Overview of AmodalCG. Our framework first determines which samples would benefit from MLLM guidance (Guidance De- cision Module). For those requiring guidance, the MLLM generates two key types of information about the occluded part of the target object: (1) the bounding box size of the full target object (Geometric Guidance Module) and (2) textual descriptions of the occluded region (Semantic Guidance Module). These are then incorporated into the completion process through a multi-scale expansion strategy, which selects the appropriate bounding box scale among the MLLM’s predictions (Inpainting Module). MaskCompletion ResultCompletion ResultMask Figure 4. Amodal completion results based on inpainting mask size. Unwanted objects are generated when the inpainting mask is substantially larger than the actual occluded region. 3.3. Estimating the Extent of Occluded Regions One type of guidance used in our framework is geometric guidance, which represents the estimated size of the full tar- get object, including its occluded regions. We first describe why this guidance is important for amodal completion and then details how the Geometric Guidance Module predicts the extent of the full target object. Geometric guidance is crucial to prevent unintended generation. A major reason existing amodal completion methods often generate undesired objects beyond the tar- get object is the use of an excessively large inpainting mask M inpaint compared to the actual occluded area. As illus- trated in Fig. 4, an inpainting mask adjusted to fit the tar- get object produces precise completions, whereas an unad- justed, overly large mask tends to generate unintended ob- jects outside the target region. Based on this observation, the Geometric Guidance Module predicts the bounding box of the full target object and redefines the inpainting mask M ∗ inpaint as the intersection between the predicted bound- ing box ˆ M bbox and the original inpainting mask M inpaint : M ∗ inpaint = ˆ M bbox ∩ M inpaint .(1) By providing the Inpainting Module with explicit guidance on how much of the object should be generated, our method effectively suppresses unnecessary object creation and pre- vents over-extension of the target object. Estimating the extent of occluded regions. This geomet- ric guidance is crucial for all samples, regardless of occlu- sion level, since even minimally occluded objects may have overly large inpainting masks (see Fig. 4 for an example). Accordingly, we adopt two different strategies based on the output of the Guidance Decision Module. If the module de- termines that MLLM guidance is unnecessary (e.g., sam- ples with minimal occlusion), the framework assumes that only minor completion is required. In this case, ˆ M bbox is obtained by slightly enlarging the modal bounding box with a fixed margin. Conversely, if MLLM guidance is deemed necessary, the framework assumes that extensive comple- tion is required, and thus ˆ M bbox is predicted by the MLLM. However, directly predicting an accurate ˆ M bbox is challeng- ing for the MLLM due to the high uncertainty inherent in occluded regions. Therefore, we jointly exploit the image generation capability of the Inpainting Module to mitigate this uncertainty. Specifically, the MLLM is instructed to predict three bounding boxes at different scales: tight, mod- erate, and coarse. Then, the Inpainting Module progres- sively evaluates each bounding box, starting from the tight- est prediction, until the object can be fully reconstructed within the prediction, as further described in Sec. 3.5. Input Prompt for the MLLM. We provide three types of information as a text prompt to the MLLM: (1) the coordi- nates of the modal bounding box, (2) the image size, and (3) 4 System Prompt Your will be provided with an image of an object, the object’s name, its visible bounding box in the format [x_min, y_min, x_max, y_max], and the size of the image as [height, width]. ..., Provide tight, moderate, and coarse bounding boxes for the full object. ... The name of the object is elephant, and its visible bounding box is [399, 80, 637, 424].The height and width of the image are [426, 640]. User: :MLLM Tight: [350, 50, 637, 424] Moderate: [320, 30, 637, 424] Coarse: [300, 20, 637, 424] Figure 5. Example of the Geometric Guidance Module predicting multi-scale bounding boxes for the full target object. Target : Horse “A dapple gray horsestanding with a rider, wearing a bridle and saddle, with muscular legs and a well-groomed coat.” “A majestic gray horsewith a strong, muscular neck and a flowing mane, wearing a bridle and saddle, standing gracefully.” Target : Plate (b) Ours(a) Without masking occluders “A white platewith crispy breaded chicken, steamed broccoli, sliced carrots, and mashed potatoes with gravy.” “A white ceramic plate with a glossy finish, subtle decorative patterns along the rim, elegant and simple design.” Figure 6. Without masking occluders, MLLMs tend to describe occluders as well, resulting in extraneous generation in the final output. We highlight the semantic category of the target object in blue and the descriptions of the occluders in red. the semantic category. For the visual prompt, we isolate the target object on a white background and highlight its modal bounding box in red, allowing the MLLM to clearly asso- ciate the coordinates provided in the textual prompt with the corresponding region in the image. Based on this input, the MLLM predicts tight, moderate, and coarse bounding boxes of the full target object. Fig. 5 illustrates our method, and the exact prompts used in the experiments are provided in Appendix Sec. 6. 3.4. Generating Detailed Textual Descriptions The Semantic Guidance Module extends beyond a cate- gory level prompt P cat by generating P long , a detailed tex- tual description specifying what should be generated in the occluded regions. This module is invoked only when the “Man sitting on steps, wearing a white shirt and khaki shorts, hands resting on knees, casual sneakers.” System Prompt Your job is to speculate the obscured part of the object inside the red box in the image and provide a Stable Diffusion prompt about the hidden part, using no more than 77 tokens. ... A part of the manin the red box is obscured by black occluders. User: :MLLM Figure 7. Example of the Semantic Guidance Module generating detailed descriptions of occluded regions. Guidance Decision Module determines that MLLM guid- ance is necessary. Below, we describe how the MLLM gen- erates descriptions of occluded regions. P = ( P long , if Semantic Guidance Module is invoked, P cat ,otherwise. Unlike existing description generation methods that fo- cus on visible parts of scenes, our focus is on generating descriptions for occluded regions of an object. However, describing the occluded parts, rather than the visible ones, presents a unique challenge. As shown in Figure 6, existing methods [2, 5, 23], which typically use visual marks to in- dicate the target object, often include descriptions of the oc- cluders when they significantly overlap with the target ob- ject, leading to the generation of unintended objects in the final outputs. In the figure, although the rider and broccoli are occluders in each image, the MLLM provides descrip- tions about them, causing their inclusion in the final outputs. Thus, in amodal completion, it is crucial to prevent occlud- ers from influencing the MLLM’s responses. Interestingly, we observe that using a visual prompt that masks the occluders effectively mitigates this issue. By re- moving the occluders’ appearance, this approach allows the MLLM to easily distinguish the occluders from the target object, thereby minimizing their influence on the MLLM’s response. It also allows the MLLM to understand the over- all context of the image by preserving the appearance of the parts outside the occluders. Figure 7 illustrates our method for generating descriptions of occluded regions. 3.5. Completion with MLLM Guidance Finally, the Inpainting Module reconstructs the appearance of the target object using the two types of guidance and out- 5 ImageModal ImageMC Diffusionpix2gestaltOursGT MaskOWAAC Figure 8. Qualitative evaluation of our method. puts the completed object ˆ I amodal along with its segmen- tation mask ˆ M amodal . We first place the target object on a gray background I bkgd and perform inpainting using the re- sized inpainting mask M ∗ inpaint and the text prompt P . Af- ter inpainting, we separate the reconstructed target object from I bkgd by obtaining the background mask ˆ M bkgd using SAM [7] and then inverting it to derive the amodal mask of the target object: ˆ M amodal = (1− ˆ M bkgd )∪ M modal . This approach is used because segmenting the background is generally easier, whereas the target object often contains complex internal details. Multi-scale expansion. As described in Sec. 3.3, the Ge- ometric Guidance Module outputs three different scales of ˆ M bbox when MLLM guidance is invoked. In such cases, the Inpainting Module leverages its image generation capa- bility to determine an appropriate mask scale among them. Specifically, the module starts from the tightest prediction and checks whether the target object can be fully recon- structed within the prediction. If the object is successfully reconstructed, the generated result is returned as the final output; otherwise, the module proceeds to the next larger scale and repeats the verification. This progressive process ensures that the target object is fully reconstructed within the predicted region while avoiding unintended generation. To verify whether the object has been fully reconstructed, we check whether the generated object touches the bound- ary of ˆ M bbox . If the generated object reaches the boundary, we regard it as incomplete; otherwise, it is considered fully reconstructed. Although our method can continue to refine the mask by incrementally enlarging it beyond the coarsest prediction, we limit the expansion process to three scales in our experiments for efficiency. 4. Experiments When evaluating amodal completion, two key aspects should be considered. The first is whether the target object is fully generated, and the second is whether the appear- ance of the object is naturally reconstructed. Following the evaluation methods of the baselines, we assess these aspects through two tasks: Amodal Segmentation and Occluded Object Recognition. We first present the performance of our method in amodal segmentation, followed by results for oc- cluded object recognition. Finally, we validate the effective- ness of each component of our method. Implementation. We use InternVL3.5-8B [20] as the Guid- ance Decision Module and GPT-4o [14] for both the Geo- metric and Semantic Guidance Modules. Stable Diffusion v2 inpainting model [17] is used for the Inpainting Mod- ule. In the Geometric Guidance Module, a 10% margin is added to the modal bounding box when MLLM guidance is not invoked. Detailed experimental settings are provided in Appendix Sec. 6. 4.1. Amodal Segmentation Evaluation Details Amodal segmentation evaluates the similarity between the segmentation mask of the completed object and the ground truth mask of the full object using mean Intersection-over-Union (mIoU). Although amodal segmentation does not consider the appearance of the recon- structed object and multiple valid ground truth masks may 6 Method COCO-ABSDS-AMP3D-A HardModerateEasyHardModerateEasyHardModerateEasy pix2gestalt64.7780.3285.9355.7486.5387.3548.8170.9278.74 MC Diffusion59.0173.1786.9057.3764.7374.5942.6664.2974.67 OWAAC51.0662.4178.7254.1064.8671.5646.7666.5275.75 Ours75.09 86.49 92.3767.09 86.60 90.2551.73 75.83 84.09 Table 1. Results on amodal segmentation by occlusion ratio. exist for the occluded regions, it allows us to assess whether the target object is accurately reconstructed without incom- pletion or overextension by comparing the similarity of the masks. We evaluate our method on three datasets: COCO- A [28], BSDS-A [28], and MP3D-A [25]. These datasets include a variety of objects commonly found in everyday life and are the most frequently used datasets for amodal segmentation. Results Table 1 shows that our method outperforms all baselines and remains the most robust under high-occlusion scenarios. Following [21], we define samples with an occlu- sion ratio above 0.5 as hard, below 0.2 as easy, and the rest as moderate. As illustrated in Figure 8, existing methods struggle with heavily occluded objects due to their inability to incorporate detailed information about the occluded re- gions. In contrast, our method generates detailed guidance about occluded regions by leveraging the rich knowledge of the MLLM, thereby significantly improves amodal seg- mentation performance compared to the baselines. Specif- ically, our method can prevent the generation of extrane- ous objects by utilizing an inpainting mask that fits the tar- get object’s size, and can also generate more natural recon- structions through progressive mask expansion guided by detailed prompts. The effectiveness of each component of our method is further explored in Sec. 4.3. 4.2. Occluded Object Recognition Evaluation Details Occluded object recognition is a classi- fication task that classifies the class of an occluded object. This task allows us to evaluate how well the appearance of the occluded object has been restored. Following the setting used in pix2gestalt, we employ CLIP [16] as the classifi- cation model and evaluate objects placed on a white back- ground. For the dataset, we utilize the Occluded and Sep- arated COCO datasets [24], which contain 80 COCO se- mantic categories. The occluded COCO dataset consists of occluded objects represented as a single segment, while the separated COCO dataset consists of occluded objects repre- sented as multiple segments, making it more challenging. Results Table 2 demonstrates that our method is highly effective in completing the appearance of occluded ob- jects compared to the baselines. As shown in the table, our method consistently outperforms existing approaches. This is because existing methods heavily rely on the prior knowl- edge of Stable Diffusion to reconstruct occluded objects Method OccludedSeparated Top 1↑Top 3↑Top 1↑Top 3↑ No completion34.0049.2621.1034.70 pix2gestalt43.3958.9731.1545.77 MC Diffusion44.7462.0734.5049.72 OWAAC40.5055.3027.8340.97 Ours45.06 62.99 40.01 56.70 Table 2. Qualitative evaluation on occluded object recognition us- ing two datasets. We report Top 1 and Top 3 accuracy (%) using CLIP as a classification model. without providing detailed guidance on what to generate for the occluded parts. In contrast, our approach leverages de- tailed information about the occluded regions generated by the MLLM to help the reconstruction of the occluded ob- ject, and our experimental results indicate that this approach is highly effective for amodal completion. Furthermore, our method performs well on the more challenging separated COCO dataset, which demonstrates that our method is also effective in difficult samples. COCO-ABSDS-AMP3D-A GCR↑GSR↑GCR↑GSR↑GCR↑GSR↑ 94.1848.7398.1735.0299.62- Table 3. Guidance Call/Skip Rate (GCR/GSR) (%) of the Guid- ance Decision Module. GSR is not reported for MP3D-A, as it does not contain samples with low occlusion (i.e., <10%). 4.3. Ablation Study Analysis of Guidance Decision Module We first evaluate whether the Guidance Decision Module appropriately in- vokes MLLM guidance when needed. The module is as- sessed in two aspects: (1) whether it correctly invokes guid- ance — Guidance Call Rate (GCR), and (2) whether it ap- propriately avoids unnecessary guidance — Guidance Skip Rate (GSR). Since ground-truth labels indicating whether guidance is required are unavailable, we approximate them by defining samples with an occlusion ratio greater than 50% as those that truly require guidance, and samples with less than 10% occlusion as those that likely do not. As shown in Tab. 3, our module accurately invokes guidance 7 Figure 9. Failure cases where the Guidance Decision Module failed to skip guidance. Method COCO-A mIoU↑ (%) SD Inpainting (w/o guidance)74.77 + geometric guidance (single scale)85.64 + semantic guidance85.60 + multi-scale expansion (Ours)86.31 Table 4. Ablation study of our framework. Each component is se- quentially added. for samples that truly require it, showing high GCR. In con- trast, GSR is relatively lower. We found that this is mainly due to cases, such as those shown in Fig. 9, where the object is truncated by the image boundary or belongs to ambiguous background regions (i.e., trees). Since the occlusion ratio is computed within the image boundary, these objects exhibit low occlusion ratios despite being incomplete, which causes the module to invoke guidance even when the ratio is low. Analysis of Geometric Guidance Module Our geometric guidance, which adjusts the inpainting mask to align with the actual size of the target object, plays a key role in per- formance improvement, as applying it alone yields strong results, as illustrated in Table 4. We attribute this to the fre- quent occurrence of substantial occluders in natural scenes, which causes the inpainting mask to become overly large and leads to extraneous generation. Additionally, we exam- ine how completion results can be refined by our multi-scale expansion strategy in Fig. 10. As illustrated, our method enables the controllable generation of objects by utilizing masks of different scales. Analysis of Semantic Guidance Module After resizing the inpainting mask to match the target object’s size, our detailed descriptions from the Semantic Guidance Module helps restore the object with a plausible pose and appear- Target objectCoarse maskModerate maskTight mask Figure 10. Completion results across different mask scales. Target : Cow ”A brown cowstanding in a field, with its head turned slightly, showing its ears and eyes.” “A brown cowstanding on all four legs, with a sturdy build, short fur, and a calm expression.” “Cow” (a)(b)(c)(d) Target : Plate “A white platewith crispy breaded chicken, steamed broccoli, sliced carrots, and mashed potatoes with gravy.” “A white ceramic plate with a glossy finish, subtle decorative patterns along the rim, elegant and simple design.” “Plate” Figure 11. (a) Image of the target object. (b) Completion results with a semantic category as a text prompt. (c) Completion results with a description generated without masking occluders. (d) Ours. ance. As shown in Fig. 11, these descriptions effectively capture the pose and visual characteristics of the hidden re- gions, enabling the generation of realistic appearances such as the cow’s patterns or a clean plate. However, the effect of semantic guidance on COCO-A is minimal, as illustrated in Tab. 4. We attribute this to the fact that COCO-A primar- ily consists of objects with limited pose variation, such as static items or natural backgrounds, and that the segmen- tation task itself does not account for appearance. In fact, our semantic guidance proves beneficial for occluded object recognition, as shown in Tab. 5. We present further analysis in Appendix Sec. 8. Method OccludedSeparated Top 1↑Top 3↑Top 1↑Top 3↑ Ours w/o semantic guidance44.5462.6539.0456.33 Ours45.06 62.99 40.01 56.70 Table 5. Effectiveness of our semantic guidance in OOR. 5. Conclusion In this paper, we proposed AmodalCG, a framework that selectively harnesses the rich real-world knowledge of MLLMs to guide amodal completion. First, the Guidance Decision Module selectively invoked MLLM guidance by assessing the level of occlusion. For samples requiring guid- ance, our framework generated two key types of guidance from MLLMs to improve the completion process. First, ge- ometric guidance provided cues on how much of the object should be reconstructed, thus preventing extraneous genera- tion. Second, semantic guidance offered detailed instruction about what should be generated. During the completion pro- cess, we exploited the strengths of both MLLMs and visual generative models through a multi-scale expansion strat- egy. Experimental results demonstrated that MLLMs can effectively enhance amodal completion, offering a promis- ing direction for integrating large multimodal reasoning into amodal completion. 8 References [1] Jiayang Ao, Yanbei Jiang, Qiuhong Ke, and Krista A Ehinger. Open-world amodal appearance completion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6490–6499, 2025. 2, 3 [2] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm’s ref- erential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 5 [3] Helisa Dhamo, Nassir Navab, and Federico Tombari. Object- driven multi-layer scene decomposition from a single image, 2019. 3 [4] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible, 2018. 3 [5] Songtao Jiang, Yan Zhang, Chenyi Zhou, Yeying Jin, Yang Feng, Jian Wu, and Zuozhu Liu. Joint visual and text prompt- ing for improved object-centric perception with multimodal large language models. arXiv preprint arXiv:2404.04514, 2024. 5 [6] Gaetano Kanizsa, Paolo Legrenzi, and Paolo Bozzi. Orga- nization in vision: Essays on gestalt perception. (No Title), 1979. 1 [7] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 3, 6 [8] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579–9589, 2024. 2, 3 [9] Hyunmin Lee and Jaesik Park.Instance-wise occlusion and depth orders in natural scenes.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21210–21221, 2022. 3 [10] Huan Ling, David Acuna, Karsten Kreis, Seung Wook Kim, and Sanja Fidler. Variational amodal object completion. In Advances in Neural Information Processing Systems, pages 16246–16257. Curran Associates, Inc., 2020. 3 [11] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 [12] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3 [13] OpenAI. Gpt-4v(ision) system card, 2023. https:// cdn.openai.com/papers/GPTV_System_Card. pdf. 2 [14] OpenAI. Gpt-4o system card, 2024. https://cdn. openai.com/gpt-4o-system-card.pdf. 2, 6 [15] Ege Ozguroglu, Ruoshi Liu, D ́ ıdac Sur ́ ıs, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick.pix2gestalt: Amodal segmentation by synthesizing wholes.In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3931–3940. IEEE Computer So- ciety, 2024. 1, 2, 3 [16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, pages 8748–8763. PMLR, 2021. 7 [17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ̈ orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 6 [18] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural In- formation Processing Systems, 35:25278–25294, 2022. 3 [19] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 [20] Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Sheng- long Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 6 [21] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Amodal com- pletion via progressive mixed context diffusion. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099–9109, 2024. 1, 2, 3, 7 [22] Xiaosheng Yan, Feigege Wang, Wenxi Liu, Yuanlong Yu, Shengfeng He, and Jia Pan. Visualizing the invisible: Oc- cluded vehicle segmentation and recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 7618–7627, 2019. 3 [23] Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. Fine-grained visual prompting. Advances in Neu- ral Information Processing Systems, 36, 2024. 5 [24] G Zhan, W Xie, and A Zisserman. A tri-layer plugin to improve occluded detection. arxiv 2022.arXiv preprint arXiv:2210.10046, 2022. 7 [25] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zis- serman. Amodal ground truth and completion in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28003–28013, 2024. 7, 2 [26] Ni Zhang, Nian Liu, Junwei Han, Kaiyuan Wan, and Ling Shao. Face de-occlusion with deep cascade guidance learn- ing. Trans. Multi., 25:3217–3229, 2023. 3 [27] Qiang Zhou, Shiyin Wang, Yitong Wang, Zilong Huang, and Xinggang Wang. Human de-occlusion: Invisible perception and recovery for humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3691–3701, 2021. 3 9 [28] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Doll ́ ar. Semantic amodal segmentation. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 1464–1472, 2017. 7 10 Integrating Multimodal Large Language Model Knowledge into Amodal Completion Supplementary Material 6. Implementation and Evaluation Details In this section, we provide the implementation details of our experiments. 6.1. Guidance Decision Module Visual Prompt We segment the visible parts of the target object and place them on a white background. Then, we crop the image with a 100-pixel margin around the target object to create the visual prompt. Text Prompt We present the exact prompt used in the Guid- ance Decision Module, which produces two outputs: (1) whether MLLM guidance is required, and (2) the semantic category of the target object. You are an expert visual annotator. You are given an image where a red bounding box highlights a target subject. Follow these instructions carefully and output only a valid JSON object. Instructions: 1. Determine amodal completion requirement: - Output "no" only if you are confident that the target object in the red box is already complete, minimally occluded, or cannot be further extended. - Otherwise, output "yes". - If you are uncertain, output "yes". 2. Identify the category name: - Output the category name of the target subject. 3. Do not include explanations, reasoning, or any text outside of the JSON. Output format: "requires extensivecompletion": "yes" | "no", "category": str Example output: "requires extensivecompletion": "yes", "category": "Bear" 6.2. Geometric Guidance Module Visual Prompt We segment the visible parts of the target object and place them on a white background. We use the full-sized image as a visual prompt, allowing the MLLM to consider the original image scale when predicting the size of the full object. Text Prompt The user prompt is described in Fig. 5. Below is the exact system prompt we used to make the MLLM predicts the size of the full target object. System Prompt You will be provided with an image of an object, the object’s name, its visible bounding box in the format [x min , y min , x max , y max ], and the size of the image as[height, width]. The bounding box is marked with a red box in the image. The object in the red box is partially obscured and your task is to estimate three bounding boxes for the entire object, including both the visible and invisible parts. Provide tight, moderate, and coarse bounding boxes for the full object in the format [x min , y min , x max , y max ]. The tight bounding box should include minimal margin around the visible parts of the object. Just provide the three bounding boxes, without explanation. 6.3. Semantic Guidance Module Visual Prompt We mark the visible parts of the target ob- ject with a bounding box and mask the occluders. Then, we crop the image with a 100-pixel margin around the tar- get object to create the visual prompt. This encourages the MLLM to focus on the surrounding regions of the target object when generating descriptions. Text Prompt The user prompt is described in Fig. 7. For the system prompt, we assign the MLLM the task of inferring the occluded parts and instruct it to describe only the target object without including any descriptions of the occluders and background. Below is the exact system prompt we used to generate descriptions of occluded regions. 1 System Prompt Your job is to speculate the obscured part of the object inside the red box in the image and provide a Stable Diffusion prompt about the hidden part, using no more than 77 tokens. Do not include the names of the occluders and descriptions of the background in the prompt, focusing solely on the object. Your response should only contain the prompt. Your response should begin with a prefix that says ‘Prompt:’. 6.4. Evaluation Details Given the inherent ambiguity of amodal completion, where multiple plausible answers can exist, previous ap- proaches [1, 21] have primarily relied on user studies rather than quantitative metrics, or have evaluated similarity with incomplete, occluded objects [1], which is not robust when objects are heavily occluded. Therefore, to reliably evaluate whether the occluded regions are properly reconstructed, we adopt amodal segmentation and occluded object recog- nition as our main quantitative evaluation metrics, follow- ing the evaluation setting of pix2gestalt [15]. Amodal seg- mentation assesses reconstruction quality by comparing the mask of the reconstructed object with human-annotated or 3D-projected ground-truth masks. Occluded object recog- nition provides a complementary evaluation by checking whether the reconstructed appearance of the occluded re- gion is consistent with the correct semantic category of the object, offering a robust measure of appearance fidelity. We clarify our evaluation protocol. Unlike other base- lines that perform multiple refinement steps through iter- ative generation, pix2gestalt completes the task in a sin- gle forward generation process without refinement. There- fore, to account for the inherent uncertainty of amodal com- pletion and to ensure a fair comparison in computational cost, we evaluate pix2gestalt by generating three samples for each input and reporting their average performance. 7. Dataset Analysis Fig. 12 shows occlusion statistics for each dataset. All datasets consist of real images, each accompanied by a sin- [0.0-0.2][0.2-0.5][0.5-1.0] Occlusion ratio 0 20 40 60 80 100 Percentage (%) COCO-A [0.0-0.2][0.2-0.5][0.5-1.0] Occlusion ratio 0 20 40 60 80 100 BSDS-A [0.0-0.2][0.2-0.5][0.5-1.0] Occlusion ratio 0 20 40 60 80 100 MP3D-A Figure 12. Occlusion percentage of the three datasets. Figure 13. Examples of the three datasets. gle annotated ground-truth mask. COCO-A and BSDS-A include human-annotated masks, while MP3D-A relies on 3D projection, which introduces slight noise (see Fig. 13). High-occlusion samples are less common in COCO-A and BSDS-A due to the difficulty of manual annotation. These datasets are widely used to evaluate real-world applicability, as they are the only real-world datasets covering numerous categories [25]. 8. Analysis on the Semantic Guidance Module In this section, we explore the effectiveness of the visual prompt used in the semantic guidance module and dis- cuss why naive prompting falls short in amodal completion. Lastly, we present failure cases of the semantic guidance module. To effectively guide amodal completion, descriptions of the occluded parts should satisfy two key conditions. First, they should exclude descriptions of the occluders, as includ- ing such descriptions may lead to the generation of unin- tended objects. Second, the descriptions must be consistent with the image context; otherwise, the target object may be unnaturally restored. Masking occluders in the visual prompt enables the MLLM to generate descriptions that meet both conditions. We believe this approach helps the MLLM clearly distin- guish the target object from the occluders, while the pre- served shape of the occluders aids in inferring a plausible pose. To demonstrate this, we compare the MLLM’s re- sponses when using visual prompts with and without oc- cluder masking. As shown in the Fig. 14, without occluder masking, the MLLM often fails to distinguish the target ob- ject from the occluders, resulting in descriptions that in- clude details about the occluders. In contrast, our visual prompt enables the MLLM to accurately distinguish the tar- get object, even when the target object and the occluders overlap significantly. Furthermore, the descriptions gener- ated using our visual prompt align well with the image con- text. 2 A dapple gray horsestanding with a rider, wearing a bridle and saddle, with muscular legs and a well-groomed coat. A majestic gray horsewith a strong, muscular neck and a flowing mane, wearing a bridle and saddle, standing gracefully. A smooth, light-colored wallwith slight imperfections and cracks, featuring a mounted green metal headboard with vertical bars. A smooth, peach-colored wallwith subtle shadows and slight imperfections. A snowy mountain with ski slopes, tall evergreen trees, and skiers in winter gear. Majestic snow-covered mountain peaks under a clear blue sky, sunlight casting soft shadows on the slopes. Butterpaper, translucent, slightly crinkled, covering a portion of a hot dog, edges slightly torn, thin and glossy texture Crumpled butter paper with translucent texture, slightly greasy, reflecting light, partially covering fried food. A white plate with crispy breaded chicken, steamed broccoli, sliced carrots, and mashed potatoes with gravy. A white ceramic platewith a glossy finish, subtle decorative patterns along the rim, elegant and simple design. A human fingerholding a bottle, partially obscured, natural skin tone, realistic detail. A human fingerwith natural skin tone, slightly bent, with visible knuckles and smooth texture. A handholding an ice cream sandwich with fingers gently wrapped around the cone. A humanhand with fingers slightly curled, natural skin tone, realistic texture, visible knuckles and fingernails. A sleek, modern towel rack with a metallic finish, holding a neatly folded towel. A sleek, chrome towel rack with a modern, minimalist design, featuring smooth, curved bars and a polished finish. Without occluder masking Ours Without occluder masking Ours A manwearing a yellow jacket, black pants, and ski boots, holding ski poles, standing on a snowy slope. Manskiing down a snowy slope, wearing a winter jacket and ski pants, holding ski poles, snow-covered mountains in the background. A white ceramic platewith a glossy finish, partially covered by vibrant green lettuce leaves and colorful vegetables. A white ceramic platewith a smooth, glossy surface and a subtle rim detail. Figure 14. A comparison of the descriptions generated by the MLLM when using visual prompts with and without occluder masking. In the descriptions, the semantic category of the target object is highlighted in bold, and descriptions of the occluders are highlighted in red. 3 Target objectOriginal inpainting maskTightModerateCoarse Figure 15. Examples of inpainting masks generated by our method. The target object is highlighted with a red box in each image. The target objects are a house in the first image, a man in the second, and a car in the third. ”A white ceramic bowl filled with sliced fruits like bananas, kiwis, and oranges, smooth glossy surface, round shape.” “Woman standing, wearing a sleeveless top and jeans, holding a microphone.” Figure 16. MLLM occasionally includes descriptions of objects that are not present in the image but are related to the image con- text. 8.1. Analysis on failure cases We observe that our method is not entirely free from the co-occurrence bias of MLLMs. As shown in Fig. 16, while our method effectively enables the MLLM to distinguish between the target object and the occluder, it occasionally includes descriptions of co-occurring objects. For instance, in the figure, the MLLM mentions sliced fruits or a micro- phone, even though they are not present in the image. 9. Analysis on the Geometric Guidance Mod- ule In this section, we further analyze our approach to pre- dicting the full target object size using the MLLM. First, we present examples of the inpainting masks generated by our method. Second, we evaluate the effectiveness of our Method COCO-ABSDS-A mIoU↑ (%)mIoU↑ (%) Original inpainting mask48.0052.02 Our inpainting mask78.9774.60 Table 6. Accuracy of our inpainting mask. method in predicting the true extent of the object. Finally, we show some failure cases of the geometric guidance mod- ule. 9.1. Examples of our inpainting mask Fig. 15 shows examples of inpainting masks generated by our method. As illustrated in the figure, our approach incor- porates three different scales of masks, each reflecting the characteristics of the target object. For instance, small in- painting masks are generated when only the man’s head is occluded, whereas larger inpainting masks are created when substantial portions of the car or the house are occluded. This demonstrates that our method produces reasonable size estimations and effectively prevents the use of excessively large inpainting masks by adjusting the mask size to match the size of the full target object. 9.2. Effectiveness of the Geometric Guidance Mod- ule To evaluate the capability of our method in estimating the target object’s size, we assess the accuracy of our resized in- painting mask derived from the predicted bounding box for 4 Tight: [430, 241, 640, 422] Moderate: [400, 220, 640, 427] Coarse: [370, 200, 640, 427] Groud Truth: [238, 167, 639, 417] Tight: [344, 376, 480, 502] Moderate: [320, 360, 480, 520] Coarse: [300, 340, 480, 540] Groud Truth: [299, 375, 479, 497] (a)(b) Figure 17. Examples of failure cases. Our method tends to avoid overly aggressive estimation, even in the case of coarse estimation. (a) Fails to predict the size of a long fence. (b) Matches the ground truth size well but still fails to predict a long fence. the full target object. We compare the similarity between the inpainting mask adjusted using the ground truth bound- ing box of the full target object and the mask adjusted by our method by computing mIoU. As shown in Tab. 6, our method leverages geometric guidance to generate inpaint- ing masks that match the target object, effectively prevent- ing the use of overly large masks. 9.3. Analysis on failure cases Although our method produces reasonable estimations in most cases, as previously discussed, it sometimes avoids ex- cessively aggressive estimation, even in the case of coarse estimation. As shown in Fig. 17, although the fences could be long, our method predicts only the size of short fences. Nevertheless, this limitation can be alleviated by incorpo- rating more scales in the multi-scale expansion, enabling further enlargement of the mask at the expense of efficiency. 10. Additional Qualitative Results of Amodal Completion In this section, we present additional qualitative results (Figs. 18 and 19) of our method in amodal completion. 5 ImageModal ImageOursImageModal ImageOurs Figure 18. Additional qualitative results of our method in amodal completion. 6 ImageModal ImageOursImageModal ImageOurs Figure 19. Additional qualitative results of our method in amodal completion. 7