Pith. sign in

REVIEW 3 major objections 6 minor 77 references

A training-free pipeline that pairs a polyp image segmenter with SAM2 reaches top video polyp segmentation results in and out of domain.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

FreeVPS pairs a per-frame polyp segmenter with frozen SAM2 tracking and two filtering modules to reduce error accumulation, improving in-domain and out-of-domain video polyp segmentation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A plausible training-free SAM2-based track-by-detect post-processor for video polyp segmentation, with solid internal ablations but soft SOTA numbers due to test-set hyperparameter tuning and no released code. the 3 major comments →

arxiv 2508.19705 v1 pith:62FKXPFT submitted 2025-08-27 cs.CV

FreeVPS: Repurposing Training-Free SAM2 for Generalizable Video Polyp Segmentation

Qiang Hu , Ying Zhou , Gepeng Ji , Nick Barnes , Qiang Li , Zhiwei Wang This is my paper
classification cs.CV
keywords video polyp segmentationSAM2training-free adaptationtrack-by-detectcolonoscopydomain generalizationmemory bank refinementtemporal consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

FreeVPS argues that video polyp segmentation should be split into two jobs: a per-frame image polyp segmenter that knows what a polyp looks like, and a tracker that knows where the same polyp is across time. The paper's claim is that SAM2, used without any training, can play the tracker if two small modules watch over its memory: intra-association filtering removes false detections by aligning and voting over a short window of frames, and inter-association refinement decides which tracked objects to keep, add, or delete before their masks re-enter SAM2's memory. In the paper's experiments, this raises the QueryNet detector from Dice 82.2 to 88.4 on in-domain SUN-SEG and from 68.0 to 77.0 on unseen PICCOLO, with similar relative gains on PolypGen, without fine-tuning SAM2. Because SAM2 is untouched, the framework promises a path to video segmentation for medical tasks where annotated video is scarce.

Core claim

On the paper's own terms, the central result is that SAM2's memory bank is not a fragile component to be avoided in medical video; it can be repurposed as the alignment and propagation engine of an automatic polyp segmenter, provided every mask entering the bank is first vetted by an image-level detector. The IAF module uses SAM2's propagation to align the detector's masks from a three-frame window to a reference frame, links the aligned masks into tracklets by pairwise IoU, and keeps one voted representative per tracklet; the IAR module then matches those voted masks against SAM2's previously tracked trajectories with the Hungarian algorithm, unions the matched pairs, admits new instances o

What carries the argument

The carrying mechanism is a memory-bank curation loop built around SAM2's propagation. SAM2's cross-frame propagation is applied twice in each of the paper's non-overlapping three-frame windows: first inside IAF, where it warps the IPS detector's masks from later frames onto the reference frame (Eq. 2), and then inside IAR, where it propagates the refined reference mask to the remaining frames. IAF constructs IoU tracklets from the aligned masks and uses a voting filter to pick the most representative segment per tracklet; IAR pairs those voted segments with SAM2's existing trajectories using Hungarian assignment, refines the memory bank with unions, births, and deaths (controlled by lambda1

Load-bearing premise

The load-bearing premise is that SAM2 can align and track the same polyp accurately across nearby frames during colonoscopy; if that alignment drifts, the voting step can freeze the wrong mask and the refinement step can feed it back into SAM2's memory, recreating the snowball effect.

What would settle it

Take a colonoscopy video with a fast pan or a polyp that leaves the field and later reappears, and compare FreeVPS's per-frame masks with expert annotations. A sharp drop in mask overlap as inter-frame motion grows, or a reappearing polyp being registered as a new object rather than the old trajectory, would show that the alignment-and-memory assumption is violated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any image polyp segmenter can be upgraded to a video segmenter without retraining; the appendix shows PolypPVT and Polyper both gain on SUN-SEG and PICCOLO when FreeVPS is added.
  • No first-frame annotation is needed, so the pipeline runs on raw, untrimmed colonoscopy videos; on the in-house LU-VPS set it outperforms object detectors, image segmenters, and video segmenters on both detection and segmentation metrics.
  • The same memory-curation recipe transfers to general video instance segmentation; with SAM2 frozen, it improves Mask R-CNN and CondInst on YouTube-VIS 2019 and 2021.
  • The modules are complementary to, not a substitute for, video training: fully fine-tuning SAM2 on SUN-SEG's training videos pushes Dice still higher on all three benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial: Because out-of-domain gains come largely from the detector stage plus temporal voting, the method's ceiling is set by image-level polyp detection; improving that detector should transfer directly to FreeVPS.
  • Editorial: IAF uses SAM2's own alignment to build the tracklets that then correct SAM2, creating a circularity worth stress-testing: under fast camera motion, alignment error could exceed detector error and the voting step could confidently pick the wrong mask.
  • Editorial: The birth/death counters give SAM2 a tracking-by-detection lifecycle; an obvious extension is identity preservation across polyp re-entry, which the current memory bank would treat as a new instance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes FreeVPS, a training-free post-processing framework that couples an image polyp segmentation (IPS) model with SAM2 for video polyp segmentation. The method operates in a track-by-detect manner: within a temporal window, the Intra-Association Filtering (IAF) module aligns IPS predictions to a reference frame via SAM2, builds tracklets by spatial IoU, and votes for consistent segments; the Inter-Association Refinement (IAR) module then associates these filtered masks with SAM2's propagated masks using the Hungarian algorithm, updating SAM2's memory bank by unioning matched masks, adding newly appeared polyps, and removing disappeared trajectories. The authors evaluate FreeVPS on SUN-SEG, PICCOLO, PolypGen, and a new untrimmed in-house dataset LU-VPS, reporting state-of-the-art Dice/IoU/TC scores without fine-tuning SAM2. The paper claims that this training-free integration balances spatiotemporal modeling and domain generalization better than both IPS and VPS baselines.

Significance. If the empirical claims hold, the contribution is practically valuable: it provides a plug-and-play, training-free way to upgrade any IPS model to video segmentation, reducing reliance on scarce annotated video data. The method is clearly described, and the internal ablations are mostly consistent. The introduction of the long-untrimmed LU-VPS dataset, if released, would also be a useful resource for the community. However, the headline results are weakened by two issues: the hyperparameters are selected on the same SUN-SEG test set used for the main comparison, and no error bars or significance tests are provided. These issues affect the strength of the central SOTA claim, although the core idea remains plausible.

major comments (3)
  1. [§4.1, Table 5, Appendix C Table 8] The default hyperparameters T=3, θ=0.5, λ1=1, λ2=3 are set 'based on the ablation studies in §4.5' and Appendix C, and the ablations are evaluated on SUN-SEG without any mention of a held-out validation split. The reported Dice of 88.4 in Table 5 and Table 8 matches the headline SUN-SEG test result in Table 1, indicating that these hyperparameters were selected using the test set. This introduces optimistic bias and makes the comparison with baselines that were not tuned on the test set unfair. Please repeat the hyperparameter selection on a validation split (e.g., a subset of SUN-SEG training videos) and report the resulting test numbers, or otherwise provide evidence that the selected values are not overfit.
  2. [Table 1, §4.2] The claimed improvement over the runner-up is small: e.g., 87.5 vs 86.6 (DV) and 88.4 vs 86.6 (DV+DI) in Dice on SUN-SEG. No error bars, confidence intervals, or significance tests are reported (the paper's own checklist item 7 answers 'No'). Since the method is deterministic given a trained IPS model, the relevant variability comes from IPS model training randomness and video-level sampling. Please provide bootstrap confidence intervals across videos or repeated IPS training runs, at least for the main comparisons in Table 1, to support the 'cutting-edge' claim.
  3. [§3.2 Eq. (2), §3.3 IAR] The pipeline assumes that SAM2's cross-frame alignment (Eq. 2) is reliable enough for IoU-based tracklet construction, and that the union operation in IAR is safe to inject into SAM2's memory bank. If SAM2 propagation drifts under fast camera motion or large appearance changes, IAF can create wrong tracklets and IAR can amplify false positives, potentially recreating the snowball effect the method is designed to prevent. The paper does not include a failure analysis or sensitivity study as a function of motion magnitude or appearance change. Adding such a diagnostic would substantially strengthen the robustness claim, especially for the long-untrimmed LU-VPS scenario.
minor comments (6)
  1. [§4.4] The metric 'F150' is used in Table 2 but is never defined. Please define it or cite a reference.
  2. [Figure 1 caption] The left part says 'IPS model (SALI [12])' but the text in §1 refers to QueryNet as the IPS model. Correct this typo.
  3. [§3.3] There is a typo: 'snowballing erros' should be 'errors'.
  4. [§4.3] 'we adopt two unseen VPS dataset' should be 'datasets'.
  5. [Appendix B / Table 2] The in-house LU-VPS dataset is used for a main experimental claim, but the data are not publicly released. Please clarify whether the dataset or annotations will be released, and if not, explain why this does not limit reproducibility of Table 2.
  6. [§4.1] The paper states code/models 'will be released after the review period.' Since the method is not yet reproducible without the code, please consider providing a reproducibility appendix with pseudo-code or an anonymous release for review.

Circularity Check

0 steps flagged

No circularity: FreeVPS is a training-free, hand-defined post-processor tested on external benchmarks; no prediction reduces by construction to its inputs.

full rationale

The paper's derivation is self-contained rather than circular. FreeVPS consists of hand-defined, training-free operations: IAF aligns IPS masks with SAM2 propagation (Eq. 2), builds tracklets by IoU thresholding, and selects representatives by a voting rule (Eq. 4); IAR associates the IAF output with SAM2 tracking via Hungarian matching (Eq. 5) and refines the memory bank with conditional union/removal rules (Sec. 3.3). None of these definitions incorporate the target metrics (Dice, IoU, MAE, TC) or the benchmark labels, and the method has no learned parameters. The evaluated configurations compare against external baselines on standard datasets, and the improvement over the QueryNet IPS baseline is measured, not assumed. The main adjacent concern is that hyperparameters T, theta, lambda1, lambda2 are selected using ablations on the SUN-SEG dataset (Sec. 4.1, Table 5, Appendix C), so the reported in-domain numbers may be optimistically biased by test-set tuning; however, that is a statistical validity concern, not a derivation-level circularity. The honest limitation in Appendix F that FreeVPS depends on a reliable IPS model also does not create a tautology: the dependence is empirical, not definitional. Therefore no circular step meets the quoting-and-reduction bar, and the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical or mathematical entities; it composes a pretrained IPS model, frozen SAM2, and hand-designed thresholds. All four hyperparameters are tuned on the SUN-SEG test benchmark, and the core behavior rests on SAM2's alignment/tracking reliability and on spatial overlap as an identity cue in colonoscopy.

free parameters (4)
  • T = 3
    Time window length selected by ablation on SUN-SEG (T=3 gives Dice 88.4 vs 86.9/87.8/87.0 for T=2/4/5) in Sec 4.5; no validation split stated.
  • theta = 0.5
    IAF pairing threshold selected by ablation on SUN-SEG (theta=0.5 gives Dice 88.4 vs 87.4/88.0/88.1 for 0.3/0.4/0.6) in Sec 4.5.
  • lambda1 = 1
    Appearance confirmation windows chosen in Appx C by ablation on SUN-SEG; lambda1=1 gives highest Dice 88.4.
  • lambda2 = 3
    Disappearance confirmation windows chosen in Appx C by ablation on SUN-SEG; lambda2=3 gives Dice 88.4 vs 83.2/87.4/88.1 for 1/2/4.
axioms (5)
  • domain assumption SAM2's memory propagation can align masks from different frames to a reference frame with sufficient spatial accuracy for IoU-based matching (Eq. 2).
    IAF's tracklet construction depends on this; if alignment drifts, tracklets and voting fail. It is supported only by ablations on one benchmark.
  • domain assumption A polyp appears as a single non-overlapping segment per frame, and after alignment the same polyp overlaps above threshold theta across consecutive frames.
    Sec 3.2 tracklet construction assumes segments from distinct frames can be paired by pairwise IoU; fast motion or deformation could violate this.
  • domain assumption The pixel-wise union of SAM2's propagated mask and IAF's voted mask is a safe refinement for matched instances (Sec 3.3).
    Union can only help if both masks are the same true polyp; if SAM2 carries an error, union can amplify it in the memory bank.
  • standard math The Hungarian algorithm with IoU cost yields correct instance association between SAM2 tracks and IPS detections.
    The algorithm is standard; the assumption is that IoU is a sufficient affinity measure, which in this domain follows from the earlier alignment assumption.
  • domain assumption Pre-trained IPS predictions are a reliable signal for new polyp appearance and disappearance; persistent IPS false positives can be filtered by voting.
    IAR adds new confirmed detections into memory; the method inherits IPS limitations, as acknowledged in Appx F.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FreeVPS: Repurposing Training-Free SAM2 for Generalizable Video Polyp Segmentation." pith.science (2026). https://pith.science/paper/62FKXPFT

@misc{pith2026250819705,
  author       = {Pith},
  title        = {Pith review of: FreeVPS: Repurposing Training-Free SAM2 for Generalizable Video Polyp Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/62FKXPFT}},
  note         = {Machine review of arXiv:2508.19705}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing video polyp segmentation (VPS) paradigms usually struggle to balance between spatiotemporal modeling and domain generalization, limiting their applicability in real clinical scenarios. To embrace this challenge, we recast the VPS task as a track-by-detect paradigm that leverages the spatial contexts captured by the image polyp segmentation (IPS) model while integrating the temporal modeling capabilities of segment anything model 2 (SAM2). However, during long-term polyp tracking in colonoscopy videos, SAM2 suffers from error accumulation, resulting in a snowball effect that compromises segmentation stability. We mitigate this issue by repurposing SAM2 as a video polyp segmenter with two training-free modules. In particular, the intra-association filtering module eliminates spatial inaccuracies originating from the detecting stage, reducing false positives. The inter-association refinement module adaptively updates the memory bank to prevent error propagation over time, enhancing temporal coherence. Both modules work synergistically to stabilize SAM2, achieving cutting-edge performance in both in-domain and out-of-domain scenarios. Furthermore, we demonstrate the robust tracking capabilities of FreeVPS in long-untrimmed colonoscopy videos, underscoring its potential reliable clinical analysis.

Figures

Figures reproduced from arXiv: 2508.19705 by Gepeng Ji, Nick Barnes, Qiang Hu, Qiang Li, Ying Zhou, Zhiwei Wang.

Figure 1
Figure 1. Figure 1: (Left) Cross-domain degradation. VPS model (SALI [12]) outperforms IPS model (SALI [12]) in ID scenarios yet lags behind in OOD scenarios. By integrating QueryNet with the proposed FreeVPS, we boost both ID and OOD performance by 7.5% and 15.4%, respectively, with less ID-to-OOD performance degradation. (Right) Snowball effect. In temporal tracking, vanilla SAM2 [22] exhibits a snowball effect, with errors… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of FreeVPS. We adopt an N(= 3)-length non-overlapping sliding window. Within each window, we first employ the Intra-Association Filtering (IAF) to discover temporally consistent predictions of the image polyp segmentation (IPS) model over multiple frames, Then, we establish the Inter-Association Refinement (IAR) module to refine the propagation results of SAM2 through consistent predictions of the… view at source ↗
Figure 3
Figure 3. Figure 3: Prediction visualization on ID (SUN-SEG [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on IAF and IAR modules: qualita [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example cases from the public polyp segmentation datasets, including six IPS datasets and [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between existing public short-trimmed VPS datasets and our in-house long [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

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

77 extracted references · 63 canonical work pages

  1. [1]

    Current and future colorectal cancer screening strategies

    Aasma Shaukat and Theodore R Levin. Current and future colorectal cancer screening strategies. Nature reviews Gastroenterology & hepatology, 19(8):521–531, 2022

    work page 2022

  2. [2]

    Reduction in colorectal cancer incidence by screening endoscopy

    Hermann Brenner, Thomas Heisser, Rafael Cardoso, and Michael Hoffmeister. Reduction in colorectal cancer incidence by screening endoscopy. Nature Reviews Gastroenterology & Hepatology, 21(2):125–133, 2024

    work page 2024

  3. [3]

    Frontiers in intelligent colonoscopy

    Ge-Peng Ji, Jingyi Liu, Peng Xu, Nick Barnes, Fahad Shahbaz Khan, Salman Khan, and Deng-Ping Fan. Frontiers in intelligent colonoscopy. arXiv preprint arXiv:2410.17241, 2024

    work page arXiv 2024

  4. [4]

    Selective feature aggregation network with area-boundary constraints for polyp segmentation

    Yuqi Fang, Cheng Chen, Yixuan Yuan, and Kai-yu Tong. Selective feature aggregation network with area-boundary constraints for polyp segmentation. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part I 22, pages 302–310. Springer, 2019

    work page 2019

  5. [5]

    Pranet: Parallel reverse attention network for polyp segmentation

    Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention , pages 263–273. Springer, 2020

    work page 2020

  6. [6]

    Non-equivalent images and pixels: Confidence-aware resampling with meta-learning mixup for polyp segmentation.Medical image analysis, 78:102394, 2022

    Xiaoqing Guo, Zhen Chen, Jun Liu, and Yixuan Yuan. Non-equivalent images and pixels: Confidence-aware resampling with meta-learning mixup for polyp segmentation.Medical image analysis, 78:102394, 2022

    work page 2022

  7. [7]

    A deep weakly semi-supervised framework for endoscopic lesion segmentation

    Yuxuan Shi, Hong Wang, Haoqin Ji, Haozhe Liu, Yuexiang Li, Nanjun He, Dong Wei, Yawen Huang, Qi Dai, Jianrong Wu, et al. A deep weakly semi-supervised framework for endoscopic lesion segmentation. Medical Image Analysis, 90:102973, 2023

    work page 2023

  8. [8]

    Querynet: A unified framework for accurate polyp segmentation and detection

    Jiaxing Chai, Zhiming Luo, Jianzhe Gao, Licun Dai, Yingxin Lai, and Shaozi Li. Querynet: A unified framework for accurate polyp segmentation and detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 544–554. Springer, 2024

    work page 2024

  9. [9]

    Video polyp segmentation: A deep learning perspective

    Ge-Peng Ji, Guobao Xiao, Yu-Cheng Chou, Deng-Ping Fan, Kai Zhao, Geng Chen, and Luc Van Gool. Video polyp segmentation: A deep learning perspective. Machine Intelligence Research, 19(6):531–549, 2022

    work page 2022

  10. [10]

    Diff-vps: Video polyp segmentation via a multi-task diffusion network with adversarial temporal reasoning

    Yingling Lu, Yijun Yang, Zhaohu Xing, Qiong Wang, and Lei Zhu. Diff-vps: Video polyp segmentation via a multi-task diffusion network with adversarial temporal reasoning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 165–175. Springer, 2024

    work page 2024

  11. [11]

    Vivim: A video vision mamba for medical video segmentation

    Yijun Yang, Zhaohu Xing, Lequan Yu, Chunwang Huang, Huazhu Fu, and Lei Zhu. Vivim: A video vision mamba for medical video segmentation. arXiv preprint arXiv:2401.14168, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Sali: Short- term alignment and long-term interaction network for colonoscopy video polyp segmentation

    Qiang Hu, Zhenyu Yi, Ying Zhou, Fang Peng, Mei Liu, Qiang Li, and Zhiwei Wang. Sali: Short- term alignment and long-term interaction network for colonoscopy video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 531–541. Springer, 2024

    work page 2024

  13. [13]

    Video instance segmentation

    Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5188–5197, 2019

    work page 2019

  14. [14]

    Sipmask: Spatial information preservation for fast image and video instance segmentation

    Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16, pages 1–18. Springer, 2020

    work page 2020

  15. [15]

    Sg-net: Spatial granularity network for one-stage video instance segmentation

    Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9816–9825, 2021. 10

    work page 2021

  16. [16]

    Crossover learning for fast online video instance segmentation

    Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In proceedings of the IEEE/CVF international conference on computer vision, pages 8043–8052, 2021

    work page 2021

  17. [17]

    Deepsort: deep convolutional networks for sorting haploid maize seeds

    Balaji Veeramani, John W Raymond, and Pritam Chanda. Deepsort: deep convolutional networks for sorting haploid maize seeds. BMC bioinformatics, 19:1–9, 2018

    work page 2018

  18. [18]

    Tracking anything with decoupled video segmentation

    Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1316–1326, 2023

    work page 2023

  19. [19]

    Strongsort: Make deepsort great again

    Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. Strongsort: Make deepsort great again. IEEE Transactions on Multimedia , 25:8725–8737, 2023

    work page 2023

  20. [20]

    A new approach to linear filtering and prediction problems

    Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. 1960

    1960

  21. [21]

    A strong baseline and batch normalization neck for deep person re-identification

    Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, and Jianyang Gu. A strong baseline and batch normalization neck for deep person re-identification. IEEE Transactions on Multimedia, 22(10):2597–2609, 2019

    work page 2019

  22. [22]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Segment anything for videos: A systematic survey

    Chunhui Zhang, Yawen Cui, Weilin Lin, Guanjie Huang, Yan Rong, Li Liu, and Shiguang Shan. Segment anything for videos: A systematic survey. arXiv preprint arXiv:2408.08315, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Segment anything in medical images and videos: Benchmark and deployment

    Jun Ma, Sumin Kim, Feifei Li, Mohammed Baharoon, Reza Asakereh, Hongwei Lyu, and Bo Wang. Segment anything in medical images and videos: Benchmark and deployment. arXiv preprint arXiv:2408.03322, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234–241. Springer, 2015

    2015

  26. [26]

    Polyp-pvt: Polyp segmentation with pyramid vision transformers

    Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021

    Pith/arXiv arXiv 2021

  27. [27]

    Polyp-mamba: Polyp segmentation with visual mamba

    Zhongxing Xu, Feilong Tang, Zhe Chen, Zheng Zhou, Weishan Wu, Yuyao Yang, Yu Liang, Jiyu Jiang, Xuyue Cai, and Jionglong Su. Polyp-mamba: Polyp segmentation with visual mamba. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 510–521. Springer, 2024

    work page 2024

  28. [28]

    Toward em- bedded detection of polyps in wce images for early diagnosis of colorectal cancer

    Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward em- bedded detection of polyps in wce images for early diagnosis of colorectal cancer. International journal of computer assisted radiology and surgery, 9:283–293, 2014

    work page 2014

  29. [29]

    Wm-dova maps for accurate polyp highlighting in colonoscopy: Val- idation vs

    Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilariño. Wm-dova maps for accurate polyp highlighting in colonoscopy: Val- idation vs. saliency maps from physicians. Computerized medical imaging and graphics , 43:99–111, 2015

    work page 2015

  30. [30]

    Automated polyp detection in colonoscopy videos using shape and context information

    Nima Tajbakhsh, Suryakanth R Gurudu, and Jianming Liang. Automated polyp detection in colonoscopy videos using shape and context information. IEEE transactions on medical imaging, 35(2):630–644, 2015

    work page 2015

  31. [31]

    A benchmark for endoluminal scene segmentation of colonoscopy images

    David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, Aaron Courville, et al. A benchmark for endoluminal scene segmentation of colonoscopy images. Journal of healthcare engineering, 2017, 2017. 11

    work page 2017

  32. [32]

    Kvasir-seg: A segmented polyp dataset

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020

    work page 2020

  33. [33]

    Benchmarking polyp segmentation methods in narrow- band imaging colonoscopy images

    Guanghui Yue, Guibin Zhuo, Siying Li, Tianwei Zhou, Jingfeng Du, Weiqing Yan, Jingwen Hou, Weide Liu, and Tianfu Wang. Benchmarking polyp segmentation methods in narrow- band imaging colonoscopy images. IEEE Journal of Biomedical and Health Informatics , 27(7):3360–3371, 2023

    work page 2023

  34. [34]

    Endoscopic polyp segmentation using a hybrid 2d/3d cnn

    Juana González-Bueno Puyal, Kanwal K Bhatia, Patrick Brandao, Omer F Ahmad, Daniel Toth, Rawen Kader, Laurence Lovat, Peter Mountney, and Danail Stoyanov. Endoscopic polyp segmentation using a hybrid 2d/3d cnn. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedin...

    work page 2020

  35. [35]

    Progressively normalized self-attention network for video polyp segmentation

    Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 142–152. Springer, 2021

    work page 2021

  36. [36]

    An embedding-unleashing video polyp segmentation framework via region linking and scale alignment

    Zhixue Fang, Xinrong Guo, Jingyin Lin, Huisi Wu, and Jing Qin. An embedding-unleashing video polyp segmentation framework via region linking and scale alignment. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 1744–1752, 2024

    work page 2024

  37. [37]

    Video object segmentation using space-time memory networks

    Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9226–9235, 2019

    work page 2019

  38. [38]

    Learning what to learn for video object segmentation

    Goutam Bhat, Felix Järemo Lawin, Martin Danelljan, Andreas Robinson, Michael Felsberg, Luc Van Gool, and Radu Timofte. Learning what to learn for video object segmentation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 777–794. Springer, 2020

    work page 2020

  39. [39]

    Learning fast and robust target models for video object segmentation

    Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7406–7415, 2020

    work page 2020

  40. [40]

    Rethinking space-time networks with improved memory coverage for efficient video object segmentation

    Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems, 34:11781–11794, 2021

    work page 2021

  41. [41]

    Video object segmentation using kernelized memory network with multiple kernels

    Hongje Seong, Junhyuk Hyun, and Euntai Kim. Video object segmentation using kernelized memory network with multiple kernels. IEEE transactions on pattern analysis and machine intelligence, 45(2):2595–2612, 2022

    work page 2022

  42. [42]

    Boosting video object segmentation via space-time correspondence learning

    Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2246–2256, 2023

    work page 2023

  43. [43]

    Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory

    Cheng-Yen Yang, Hsiang-Wei Huang, Wenhao Chai, Zhongyu Jiang, and Jenq-Neng Hwang. Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory. arXiv preprint arXiv:2411.11922, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree

    Shuangrui Ding, Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Yuwei Guo, Dahua Lin, and Jiaqi Wang. Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree. arXiv preprint arXiv:2410.16268, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    A distractor-aware memory for visual object tracking with sam2

    Jovana Videnovic, Alan Lukezic, and Matej Kristan. A distractor-aware memory for visual object tracking with sam2. arXiv preprint arXiv:2411.17576, 2024. 12

    Pith/arXiv arXiv 2024

  46. [46]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213–229. Springer, 2020

    2020

  47. [47]

    The hungarian method for the assignment problem

    Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83–97, 1955

    1955

  48. [48]

    Piccolo white-light and narrow-band imaging colonoscopic dataset: A performance comparative of models and datasets

    Luisa F Sánchez-Peralta, J Blas Pagador, Artzai Picón, Ángel José Calderón, Francisco Polo, Nagore Andraka, Roberto Bilbao, Ben Glover, Cristina L Saratxaga, and Francisco M Sánchez- Margallo. Piccolo white-light and narrow-band imaging colonoscopic dataset: A performance comparative of models and datasets. Applied Sciences, 10(23):8501, 2020

    work page 2020

  49. [49]

    A multi- centre polyp detection and segmentation dataset for generalisability assessment

    Sharib Ali, Debesh Jha, Noha Ghatwary, Stefano Realdon, Renato Cannizzaro, Osama E Salem, Dominique Lamarque, Christian Daul, Michael A Riegler, Kim V Anonsen, et al. A multi- centre polyp detection and segmentation dataset for generalisability assessment. Scientific Data, 10(1):75, 2023

    work page 2023

  50. [50]

    Polyper: Boundary sensitive polyp segmentation

    Hao Shao, Yang Zhang, and Qibin Hou. Polyper: Boundary sensitive polyp segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4731–4739, 2024

    work page 2024

  51. [51]

    Efficient semantic video segmentation with per-frame inference

    Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16 , pages 352–368. Springer, 2020

    work page 2020

  52. [52]

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

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441–29454. PMLR, 2023

    work page 2023

  53. [53]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    2019

  54. [54]

    Ultralytics yolo11, 2024

    Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024

    2024

  55. [55]

    Relation detr: Exploring explicit position relation prior for object detection

    Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, Badong Chen, and Xuguang Lan. Relation detr: Exploring explicit position relation prior for object detection. In European Conference on Computer Vision, pages 89–105. Springer, 2024

    work page 2024

  56. [56]

    Tsdetector: Temporal–spatial self-correction collaborative learning for colonoscopy video detection

    Kai-Ni Wang, Haolin Wang, Guang-Quan Zhou, Yangang Wang, Ling Yang, Yang Chen, and Shuo Li. Tsdetector: Temporal–spatial self-correction collaborative learning for colonoscopy video detection. Medical Image Analysis, 100:103384, 2025

    work page 2025

  57. [57]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022

    work page 2022

  58. [58]

    Ctvis: Consistent training for online video instance segmentation

    Kaining Ying, Qing Zhong, Weian Mao, Zhenhua Wang, Hao Chen, Lin Yuanbo Wu, Yifan Liu, Chengxiang Fan, Yunzhi Zhuge, and Chunhua Shen. Ctvis: Consistent training for online video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 899–908, 2023

    work page 2023

  59. [59]

    Ning Xu, Linjie Yang, Jianchao Yang, Dingcheng Yue, Yuchen Fan, Yuchen Liang, and Thomas S. Huang. Youtube-vis dataset 2021 version.https://youtube-vos.org/dataset/ vis, 2021

    work page 2021

  60. [60]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017

    work page 2017

  61. [61]

    #IMG" and “#VID

    Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 282–298. Springer, 2020. 13 A Public Datasets In this work, we mainly used nine public polyp segmentation datasets, including six image polyp segmen...

    work page 2020

  62. [62]

    Our experimental results also support our claim

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: Our abstract and introduction clearly state main contributions made in the paper. Our experimental results also support our claim. Guidelines: • The answer NA means that the abstract and introductio...

    work page

  63. [63]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We have a dedicated section to discuss the limitations of our work. Please refer to Appx. F. Guidelines: • The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those a...

    work page

  64. [64]

    Guidelines: • The answer NA means that the paper does not include theoretical results

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? 17 Answer: [NA] Justification: The paper does not include theoretical results. Guidelines: • The answer NA means that the paper does not include theoretical results. • All the theorems, formulas, and p...

    work page

  65. [65]

    The detailed training setting and experimental setup are given in Section 4.2

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: The proposed model is desc...

    work page

  66. [66]

    The detailed descriptions of these datasets are presented in Section 4.1 and Appx

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: All of the datasets involved in this work are open-source and accessible, except one in-house dataset for e...

    work page

  67. [67]

    The trained model and code will be released after the review period

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: The detailed experiment setting is described in Section 4.2. The trained model and code will be released afte...

    work page

  68. [68]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [No] Justification: We follow the common practice in prior works and report the performance number on the standard benchmarks. Guidelines: • The answer...

    work page

  69. [69]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The information on computer resources is provided in Section 4.2. Guidelines: • The answer NA means tha...

    work page

  70. [70]

    Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: Research conducted in the paper conforms with the NeurIPS Code of Ethics. Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code...

    work page

  71. [71]

    Please refer to Appx

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: We have a dedicated section to discuss the societal impacts of our work. Please refer to Appx. G. 20 Guidelines: • The answer NA means that there is no societal impact of the work perfo...

    work page

  72. [72]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: Our work does not pose such risks to the best of our knowledge. Guidelines: • The answer NA mea...

    work page

  73. [73]

    Guidelines: • The answer NA means that the paper does not use existing assets

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We properly credited the creators or original owners of assets (e.g., code, data, models), used ...

    work page

  74. [74]

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: Our new assets introduced in the paper are well documented in the Section 4.1 and Appx. B. Guidelines: • The answer NA means that the paper does not release new assets. • Researchers should commun...

    work page

  75. [75]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [NA] Justification: Our research does not involve crowdsourcing nor resea...

    work page

  76. [76]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

    work page

  77. [77]

    Answer: [NA] Justification: Our research does not involve LLMs as any important, original, or non-standard components

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, decla...

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.