pith. machine review for the scientific record. sign in

arxiv: 2604.20748 · v1 · submitted 2026-04-22 · 💻 cs.CV

Recognition: unknown

Amodal SAM: A Unified Amodal Segmentation Framework with Generalization

Bo Zhang , Zhuotao Tian , Xin Tao , Songlin Tang , Jun Yu , Wenjie Pei
Authors on Pith no claims yet

Pith reviewed 2026-05-10 01:01 UTC · model grok-4.3

classification 💻 cs.CV
keywords amodal segmentationoccluded object completionSegment Anything Modelsynthetic occlusion dataadapter modulegeneralization to novel categoriesimage and video segmentation
0
0 comments X

The pith

Amodal SAM adds a lightweight adapter and synthetic occlusion data to SAM so the model can predict complete object shapes including hidden parts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Amodal SAM as a way to extend the Segment Anything Model to amodal segmentation, which requires filling in occluded regions of objects. It keeps SAM's backbone frozen and adds a Spatial Completion Adapter for reconstruction, uses a Target-Aware Occlusion Synthesis pipeline to create varied training examples from limited annotations, and applies new objectives for consistency and topology. Experiments show the resulting system reaches state-of-the-art accuracy on existing benchmarks while also working on object categories and scenes absent from training. The work targets practical use in environments where objects are routinely partly obscured.

Core claim

Amodal SAM extends SAM for both image and video amodal segmentation by inserting a lightweight Spatial Completion Adapter that reconstructs occluded regions while freezing the original SAM features. Training relies on a Target-Aware Occlusion Synthesis pipeline that generates diverse synthetic occlusions and on learning objectives that enforce regional consistency and topological regularization. This produces state-of-the-art results on standard benchmarks together with strong generalization to novel object categories and unseen contexts.

What carries the argument

The Spatial Completion Adapter is a lightweight module added to frozen SAM that reconstructs occluded object regions; it is trained with the Target-Aware Occlusion Synthesis pipeline for data creation and with consistency and regularization objectives.

Load-bearing premise

The synthetic occlusions generated by the Target-Aware Occlusion Synthesis pipeline have statistics and difficulty close enough to real-world cases that the adapter generalizes beyond the training distribution.

What would settle it

A direct comparison on real images of novel object categories with natural occlusions would falsify the generalization claim if Amodal SAM performs no better than unmodified SAM.

Figures

Figures reproduced from arXiv: 2604.20748 by Bo Zhang, Jun Yu, Songlin Tang, Wenjie Pei, Xin Tao, Zhuotao Tian.

Figure 1
Figure 1. Figure 1: This figure shows the comparisons between our Amodal SAM and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall structure of the proposed Amodal Segment Anything Model(SAM) includes three key aspects for the adaptation from SAM to Amodal [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Spatial Completion Adapter combines the image features and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The illustration of the proposed Target-Aware Occlusion Synthesis (TAOS) pipeline. Initially, by randomly selecting masks and superimposing them [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Examples of the datasets we constructed. For each example showing the original image, the image after the occlusion was added, the amodal mask, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The discriminator uses the predictions from Amodal SAM as negative [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The image encoder of SAM-2 consists of four stages with distinct [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The occlusion mask delineates the occluded region within the amodal [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: This figure presents the qualitative results of VRSP, C2F-Seg, and our method on the KINS and COCOA datasets. “GT” represents the ground truth, [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: More qualitative results of Amodal SAM. ”GT” represents the ground truth and the ”Predict Mask” is the result of Amodal SAM prediction. [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: This figure illustrates the viability of transitioning our method to SAM-2. [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

Amodal segmentation is a challenging task that aims to predict the complete geometric shape of objects, including their occluded regions. Although existing methods primarily focus on amodal segmentation within the training domain, these approaches often lack the generalization capacity to extend effectively to novel object categories and unseen contexts. This paper introduces Amodal SAM, a unified framework that leverages SAM (Segment Anything Model) for both amodal image and amodal video segmentation. Amodal SAM preserves the powerful generalization ability of SAM while extending its inherent capabilities to the amodal segmentation task. The improvements lie in three aspects: (1) a lightweight Spatial Completion Adapter that enables occluded region reconstruction, (2) a Target-Aware Occlusion Synthesis (TAOS) pipeline that addresses the scarcity of amodal annotations by generating diverse synthetic training data, and (3) novel learning objectives that enforce regional consistency and topological regularization. Extensive experiments demonstrate that Amodal SAM achieves state-of-the-art performance on standard benchmarks, while simultaneously exhibiting robust generalization to novel scenarios. We anticipate that this research will advance the field toward practical amodal segmentation systems capable of operating effectively in unconstrained real-world environments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces Amodal SAM, a unified framework that adapts the Segment Anything Model (SAM) for amodal segmentation of both images and videos. It adds a lightweight Spatial Completion Adapter for reconstructing occluded regions, a Target-Aware Occlusion Synthesis (TAOS) pipeline to generate synthetic training data addressing annotation scarcity, and new learning objectives for regional consistency and topological regularization. The central claims are state-of-the-art performance on standard benchmarks together with robust generalization to novel object categories and unseen contexts.

Significance. If the empirical results hold, the work would meaningfully advance amodal segmentation by preserving SAM's strong zero-shot generalization while using synthetic data to overcome the lack of amodal annotations. The unified image-video treatment and focus on practical unconstrained settings are notable strengths that could influence downstream applications in robotics and scene understanding.

major comments (3)
  1. [§4.2] §4.2 (TAOS pipeline): The description of target-aware occlusion synthesis provides no quantitative comparison of key statistics (occlusion ratio histograms, boundary complexity, number of overlapping instances, or topological features) between TAOS-generated data and real amodal datasets such as KINS or COCOA. Because the generalization results in §5.3–5.4 rest on the assumption that synthetic occlusions sufficiently match real-world distributions, this omission directly threatens the validity of the adapter's learned behavior and the headline generalization claim.
  2. [Table 1, §5.1] Table 1 and §5.1: The SOTA performance numbers are reported without error bars, multiple random seeds, or ablations that isolate the contribution of TAOS from the adapter and the new objectives. Given the synthetic nature of the training data, the absence of these controls makes it impossible to determine whether the reported gains are robust or could be artifacts of the particular TAOS hyperparameter choices.
  3. [§3.3] §3.3 (Learning Objectives): The topological regularization term is motivated but its concrete effect on failure modes (e.g., thin structures or multiply occluded objects) is not analyzed; an ablation showing how removing this term changes performance on the novel-scenario test sets would be required to substantiate that it contributes to the claimed generalization rather than merely regularizing the synthetic training distribution.
minor comments (2)
  1. [Figure 3] Figure 3: The caption does not clearly indicate whether the visualized occlusions are real or TAOS-generated, nor does it label the specific failure modes being highlighted.
  2. [Abstract] The abstract states that experiments demonstrate SOTA results and generalization, yet the main text should ensure every quantitative claim is accompanied by the corresponding table or figure reference in the same paragraph.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, indicating the revisions we will incorporate to strengthen the manuscript's claims regarding data fidelity, experimental robustness, and component contributions.

read point-by-point responses

  1. Referee: [§4.2] §4.2 (TAOS pipeline): The description of target-aware occlusion synthesis provides no quantitative comparison of key statistics (occlusion ratio histograms, boundary complexity, number of overlapping instances, or topological features) between TAOS-generated data and real amodal datasets such as KINS or COCOA. Because the generalization results in §5.3–5.4 rest on the assumption that synthetic occlusions sufficiently match real-world distributions, this omission directly threatens the validity of the adapter's learned behavior and the headline generalization claim.

    Authors: We agree that a direct quantitative comparison of occlusion statistics would strengthen the justification for using TAOS-generated data to support generalization. In the revised manuscript, we will expand §4.2 with a new analysis subsection and accompanying figure that reports occlusion ratio histograms, boundary complexity metrics, number of overlapping instances, and topological features, comparing TAOS outputs directly against KINS and COCOA. This addition will explicitly validate the distributional match and bolster the generalization results in §5.3–5.4. revision: yes

  2. Referee: [Table 1, §5.1] Table 1 and §5.1: The SOTA performance numbers are reported without error bars, multiple random seeds, or ablations that isolate the contribution of TAOS from the adapter and the new objectives. Given the synthetic nature of the training data, the absence of these controls makes it impossible to determine whether the reported gains are robust or could be artifacts of the particular TAOS hyperparameter choices.

    Authors: We concur that the lack of error bars, multi-seed statistics, and component-isolating ablations limits the ability to assess robustness, particularly with synthetic training data. In the revised version, we will update Table 1 and §5.1 to report mean and standard deviation over multiple random seeds (at least three runs) and insert new ablation tables that separately quantify the contributions of TAOS, the Spatial Completion Adapter, and the learning objectives. These controls will clarify that the SOTA gains are not artifacts of specific hyperparameter settings. revision: yes

  3. Referee: [§3.3] §3.3 (Learning Objectives): The topological regularization term is motivated but its concrete effect on failure modes (e.g., thin structures or multiply occluded objects) is not analyzed; an ablation showing how removing this term changes performance on the novel-scenario test sets would be required to substantiate that it contributes to the claimed generalization rather than merely regularizing the synthetic training distribution.

    Authors: We recognize that demonstrating the specific impact of the topological regularization term on generalization to novel scenarios would better substantiate its role beyond synthetic-data regularization. We will add a targeted ablation study (in §5.3 or a new subsection) that removes this term and evaluates performance changes on the novel-scenario test sets, with particular attention to failure modes involving thin structures and multiply occluded objects. This will directly address the concern and clarify the term's contribution to the overall generalization claims. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical framework with independent benchmark validation

full rationale

The paper introduces Amodal SAM as a composite system (frozen SAM + lightweight adapter + TAOS synthesis + new objectives) whose performance claims rest on external benchmark results and held-out novel-scenario tests rather than any self-referential equation or fitted parameter renamed as a prediction. No mathematical derivations appear that equate outputs to inputs by construction, and no load-bearing self-citations or uniqueness theorems are invoked. The derivation chain is therefore self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 2 invented entities

The framework rests on the assumption that SAM features transfer to amodal completion and that synthetic occlusions suffice for training. It introduces two new modules whose parameters are learned from data and one new data-generation procedure.

free parameters (2)
  • Spatial Completion Adapter weights
    Trainable parameters of the lightweight adapter fitted to amodal data.
  • TAOS generation hyperparameters
    Parameters controlling how occlusions are synthesized; chosen or tuned during pipeline design.
axioms (1)
  • domain assumption SAM's pre-trained image encoder features contain sufficient information for occluded region reconstruction when augmented by a small adapter.
    Invoked by the decision to keep SAM frozen and add only a lightweight adapter.
invented entities (2)
  • Spatial Completion Adapter no independent evidence
    purpose: Module that reconstructs occluded regions while preserving SAM's generalization.
    New architectural component introduced by the paper.
  • Target-Aware Occlusion Synthesis (TAOS) pipeline no independent evidence
    purpose: Procedure to generate diverse synthetic amodal training examples.
    New data synthesis method introduced by the paper.

pith-pipeline@v0.9.0 · 5506 in / 1394 out tokens · 68139 ms · 2026-05-10T01:01:17.489601+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    Segment Anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W. Lo, P. Doll ´ar, and R. B. Girshick, “Segment anything,” inIEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. IEEE, 2023, pp. 3992–4003. [Online]. Available: https://doi.org/10.1109/ICCV51070.2023.00371

    work page doi:10.1109/iccv51070.2023.00371 2023

  2. [2]

    Semantic amodal segmentation,

    Y . Zhu, Y . Tian, D. N. Metaxas, and P. Doll ´ar, “Semantic amodal segmentation,” in2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. IEEE Computer Society, 2017, pp. 3001–3009. [Online]. Available: https://doi.org/10.1109/CVPR.2017.320

    work page doi:10.1109/cvpr.2017.320 2017

  3. [3]

    URL https://doi.org/10.1007/ 978-3-319-46475-6_5

    K. Li and J. Malik, “Amodal instance segmentation,” inComputer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II, ser. Lecture Notes in Computer Science, B. Leibe, J. Matas, N. Sebe, and M. Welling, Eds., vol. 9906. Springer, 2016, pp. 677–693. [Online]. Available: https://doi.org/10.1007/...

    work page doi:10.1007/978-3-319-46475-6 2016

  4. [4]

    A weakly supervised amodal segmenter with boundary uncertainty estimation,

    K. Nguyen and S. Todorovic, “A weakly supervised amodal segmenter with boundary uncertainty estimation,” inICCV, 2021

    2021

  5. [5]

    Shapeformer: Shape prior visible-to-amodal transformer-based amodal instance segmentation,

    M. Tran, W. Bounsavy, K. V o, A. Nguyen, T. Nguyen, and N. Le, “Shapeformer: Shape prior visible-to-amodal transformer-based amodal instance segmentation,” inIJCNN, 2024

    2024

  6. [6]

    Aisformer: Amodal instance segmentation with transformer,

    M. Q. Tran, K. V o, K. Yamazaki, A. A. F. Fernandes, M. Kidd, and N. Le, “Aisformer: Amodal instance segmentation with transformer,” in BMVC, 2022

    2022

  7. [7]

    Amodal instance segmen- tation with KINS dataset,

    L. Qi, L. Jiang, S. Liu, X. Shen, and J. Jia, “Amodal instance segmen- tation with KINS dataset,” inCVPR, 2019

    2019

  8. [8]

    Amodal ground truth and completion in the wild,

    G. Zhan, C. Zheng, W. Xie, and A. Zisserman, “Amodal ground truth and completion in the wild,” inCVPR, 2024

    2024

  9. [9]

    Learning to see the invisible: End-to-end trainable amodal instance segmentation,

    P. Follmann, R. K ¨onig, P. H ¨artinger, M. Klostermann, and T. B ¨ottger, “Learning to see the invisible: End-to-end trainable amodal instance segmentation,” inWACV, 2019

    2019

  10. [10]

    Learning semantics-aware distance map with semantics layering network for amodal instance segmentation,

    Z. Zhang, A. Chen, L. Xie, J. Yu, and S. Gao, “Learning semantics-aware distance map with semantics layering network for amodal instance segmentation,” inProceedings of the 27th ACM International Conference on Multimedia, ser. MM ’19. ACM, Oct. 2019, p. 2124–2132. [Online]. Available: http://dx.doi.org/10.1145/ 3343031.3350911

    work page arXiv 2019

  11. [11]

    U-Net: convolutional networks for biomedical image segmentation , booktitle =

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, ser. Lecture Notes in Computer Science, N. Navab, J. Hornegger, W. M. W. III, and...

    work page doi:10.1007/978-3-319-24574-4 2015

  12. [12]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” inCVPR, 2015

    2015

  13. [13]

    Mask R-CNN,

    K. He, G. Gkioxari, P. Doll ´ar, and R. B. Girshick, “Mask R-CNN,” in ICCV, 2017

    2017

  14. [14]

    Sam fails to seg- ment anything?–sam-adapter: Adapting sam in underperformed scenes: Camou- flage, shadow, and more.arXiv:2304.09148, 2023

    T. Chen, L. Zhu, C. Ding, R. Cao, Y . Wang, Z. Li, L. Sun, P. Mao, and Y . Zang, “SAM fails to segment anything? - sam-adapter: Adapting SAM in underperformed scenes: Camouflage, shadow, and more,”CoRR, vol. abs/2304.09148, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2304.09148

    work page doi:10.48550/arxiv.2304.09148 2023

  15. [15]

    Semantic-sam: Segment and recognize anything at any granularity

    F. Li, H. Zhang, P. Sun, X. Zou, S. Liu, J. Yang, C. Li, L. Zhang, and J. Gao, “Semantic-sam: Segment and recognize anything at any granularity,”CoRR, vol. abs/2307.04767, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2307.04767

    work page doi:10.48550/arxiv.2307.04767 2023

  16. [16]

    Segment anything in medical images,

    J. Ma and B. Wang, “Segment anything in medical images,”CoRR, vol. abs/2304.12306, 2023. [Online]. Available: https://doi.org/10.48550/ arXiv.2304.12306

    work page arXiv 2023

  17. [17]

    Segment anything in high quality,

    L. Ke, M. Ye, M. Danelljan, Y . Liu, Y . Tai, C. Tang, and F. Yu, “Segment anything in high quality,” inNeurIPS, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., 2023

    2023

  18. [18]

    Parameter-efficient transfer learning for NLP,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. de Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for NLP,” inICML, K. Chaudhuri and R. Salakhutdinov, Eds., 2019

    2019

  19. [19]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” inICLR, 2022. IEEE TRANSACTIONS ON IMAGE PROCESSING 12

    2022

  20. [20]

    Vision transformer adapter for dense predictions,

    Z. Chen, Y . Duan, W. Wang, J. He, T. Lu, J. Dai, and Y . Qiao, “Vision transformer adapter for dense predictions,” inICLR, 2023

    2023

  21. [21]

    SAM 2: Segment Anything in Images and Videos

    N. Ravi, V . Gabeur, Y . Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C. Wu, R. B. Girshick, P. Doll ´ar, and C. Feichtenhofer, “SAM 2: Segment anything in images and videos,”CoRR, vol. abs/2408.00714,

    work page internal anchor Pith review arXiv

  22. [22]

    SAM 2: Segment Anything in Images and Videos

    [Online]. Available: https://doi.org/10.48550/arXiv.2408.00714

    work page internal anchor Pith review doi:10.48550/arxiv.2408.00714

  23. [23]

    Unsupervised deep metric learning with transformed attention consistency and contrastive clustering loss,

    Y . Li, S. Kan, and Z. He, “Unsupervised deep metric learning with transformed attention consistency and contrastive clustering loss,” in ECCV, A. Vedaldi, H. Bischof, T. Brox, and J. Frahm, Eds., 2020

    2020

  24. [24]

    Tokens-to-token vit: Training vision transformers from scratch on imagenet,

    L. Yuan, Y . Chen, T. Wang, W. Yu, Y . Shi, Z. Jiang, F. E. H. Tay, J. Feng, and S. Yan, “Tokens-to-token vit: Training vision transformers from scratch on imagenet,” inICCV, 2021

    2021

  25. [25]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inICCV, 2021

    2021

  26. [26]

    Free-form image inpainting with gated convolution,

    J. Yu, Z. Lin, J. Yang, X. Shen, X. Lu, and T. S. Huang, “Free-form image inpainting with gated convolution,” inICCV, 2019

    2019

  27. [27]

    Identification and rejection of scattered neutrons in agata,

    M. S ¸enyi˘git, A. Atac ¸, S. Akkoyun, A. Kas ¸kas ¸, D. Bazzacco, J. Nyberg, F. Recchia, S. Brambilla, F. Camera, F. Crespi, E. Farnea, A. Giaz, A. Gottardo, R. Kempley, J. Ljungvall, D. Mengoni, C. Michelagnoli, B. Million, M. Palacz, L. Pellegri, S. Riboldi, E. S ¸ahin, P. S ¨oderstr¨om, and J. Valiente Dobon, “Identification and rejection of scattered...

    work page doi:10.1016/j.nima.2013.09.035 2014

  28. [28]

    Per-pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. G. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmentation,” inNeurIPS, M. Ranzato, A. Beygelzimer, Y . N. Dauphin, P. Liang, and J. W. Vaughan, Eds., 2021

    2021

  29. [29]

    Generative adversarial networks,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville, and Y . Bengio, “Generative adversarial networks,”Commun. ACM, vol. 63, no. 11, pp. 139–144, 2020. [Online]. Available: https://doi.org/10.1145/3422622

    work page doi:10.1145/3422622 2020

  30. [30]

    Hiera: A hierarchical vision transformer without the bells- and-whistles,

    C. Ryali, Y . Hu, D. Bolya, C. Wei, H. Fan, P. Huang, V . Aggarwal, A. Chowdhury, O. Poursaeed, J. Hoffman, J. Malik, Y . Li, and C. Fe- ichtenhofer, “Hiera: A hierarchical vision transformer without the bells- and-whistles,” inICML, A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, Eds., 2023

    2023

  31. [31]

    Dn-splatter: Depth and normal priors for gaussian splatting and meshing

    P. Follmann, R. K ¨onig, P. H ¨artinger, M. Klostermann, and T. B ¨ottger, “Learning to see the invisible: End-to-end trainable amodal instance segmentation,” inIEEE Winter Conference on Applications of Computer Vision, WACV 2019, Waikoloa Village, HI, USA, January 7-11, 2019. IEEE, 2019, pp. 1328–1336. [Online]. Available: https://doi.org/10.1109/W ACV ....

    work page doi:10.1109/w 2019

  32. [32]

    Unsupervised object learning via common fate,

    M. Tangemann, S. Schneider, J. von K ¨ugelgen, F. Locatello, P. V . Gehler, T. Brox, M. K ¨ummerer, M. Bethge, and B. Sch ¨olkopf, “Unsupervised object learning via common fate,” inConference on Causal Learning and Reasoning, CLeaR 2023, 11-14 April 2023, Amazon Development Center , T ¨ubingen, Germany, April 11-14, 2023, ser. Proceedings of Machine Learn...

    2023

  33. [33]

    Coarse- to-fine amodal segmentation with shape prior,

    J. Gao, X. Qian, Y . Wang, T. Xiao, T. He, Z. Zhang, and Y . Fu, “Coarse- to-fine amodal segmentation with shape prior,” inICCV, 2023

    2023

  34. [34]

    Microsoft COCO: common objects in context,

    T. Lin, M. Maire, S. J. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft COCO: common objects in context,” inECCV, D. J. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, Eds., 2014

    2014

  35. [35]

    Self-supervised scene de-occlusion,

    X. Zhan, X. Pan, B. Dai, Z. Liu, D. Lin, and C. C. Loy, “Self-supervised scene de-occlusion,” inCVPR, 2020

    2020

  36. [36]

    Amodal segmentation based on visible region segmentation and shape prior,

    Y . Xiao, Y . Xu, Z. Zhong, W. Luo, J. Li, and S. Gao, “Amodal segmentation based on visible region segmentation and shape prior,” inAAAI, 2021

    2021

  37. [37]

    PLUG: revisiting amodal segmentation with foundation model and hierarchical focus,

    Z. Liu, L. Qiao, X. Chu, and T. Jiang, “PLUG: revisiting amodal segmentation with foundation model and hierarchical focus,”CoRR, vol. abs/2405.16094, 2024. [Online]. Available: https://doi.org/10. 48550/arXiv.2405.16094

    work page arXiv 2024

  38. [38]

    categories

    E. Ozguroglu, R. Liu, D. Sur ´ıs, D. Chen, A. Dave, P. Tokmakov, and C. V ondrick, “pix2gestalt: Amodal segmentation by synthesizing wholes,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024. IEEE, 2024, pp. 3931–3940. [Online]. Available: https://doi.org/10.1109/ CVPR52733.2024.00377

    work page arXiv 2024

  39. [39]

    Towards efficient foundation model for zero-shot amodal segmentation,

    Z. Liu, L. Qiao, X. Chu, L. Ma, and T. Jiang, “Towards efficient foundation model for zero-shot amodal segmentation,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025. Computer Vision Foundation / IEEE, 2025, pp. 20 254–20 264. [Online]. Available: https://openaccess.thecvf.com/content/CVPR...

    2025

  40. [40]

    Self-supervised amodal video object segmentation,

    J. Yao, Y . Hong, C. Wang, T. Xiao, T. He, F. Locatello, D. P. Wipf, Y . Fu, and Z. Zhang, “Self-supervised amodal video object segmentation,” inAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, S. Koyejo, S. Mohamed, A....

    2022