- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This work handles the occlusion issues in unsupervised video object centric learning. The proposed method, iSEE, detects occlusion using the attention mass of a slot by comparing with its past, decomposes the slot vector into appearance and position, and moves the held position with a walker that is trained via when the object re-appears.
Critical strengths
- The diagnosis is clear and intuitive.
- It technique TEN is simple yet effective and parameter-free.
- Walker's training is elegantly self-supervised on appearance.
Critical weaknesses
- Requires stronger real world video evidence. The experiments were mainly conducted on synthetic and relative toy real-world videos. Please report the results on YouTube-VIS, a large-scale real-world video dataset, and/or OVIS, an occlusion-heavy real-world benchmark.
- Evaluation choices seem to favour the authors's method. Static occlusions are capped, IoU thresholds are limited and RAM is reported without its official post-processing. All these choices are argued. So please also report the uncapped, unlimited results so that readers can see how much they matter together.
- Most of the gains comes from the walker, which is RAM's mechanism. The key claim, iSEE identifies "two properties", looks stronger if it stated that NoPE is a precondition for the split.
Questions for authors
The authors list a fixed slot count per dataset as a limitation. Slot pruning methods, for example SOLV (Aydemir 2023) and DIAS (Zhao 2025), address exactly this. Please at least discuss in the paper whether they can be combined with TEN. The reviewer wonders: a held slot has low evidence, a signal pruning methods also use to remove redundant slots. So will pruning delete the slots TEN is trying to keep?
Reference:
- Aydemir et al. Self-Supervised Object-Centric Learning for Videos. NeurIPS 2023.
- Zhao et al. Slot Attention with Re-initialization and Self-distillation. ACM MM 2025.
2 comments
Thank you for the reply. Most of my concerns have been addressed. I keep my positive rating.
Response to Reviewer 6V7M
Thank you for the constructive feedback and the concrete suggestions. We are glad that you find the diagnosis clear, TEN simple and effective, and the self-supervised training of the walker elegant.
(a) Could you report results on a large-scale real-world benchmark, YouTube-VIS and/or OVIS?
We appreciate the suggestion. We trained SlotContrast and iSEE from scratch on OVIS and report re-identification after full occlusion in the general response (G3). iSEE has the highest SAME in every displacement range, 6.5 to 9.4 points above SlotContrast. OVIS and KITTI have no amodal masks, so we also ran a blinded user study on clips from both datasets. Users preferred iSEE in 42.7% of the comparisons and SlotContrast in 15.7%.
In short, the gain of iSEE shines on occlusion-heavy real-world video.
(b) Could you also report the uncapped results, at all IoU thresholds, and with RAM's official post-processing?
We now report every scorable occlusion of all 1,371 test clips per camera, with RAM both raw and with its published post-processing. Please refer to the general response (G1). We have updated the paper appendix with this table, Appendix C.6 (Table S8).
As reported in Appendix B.4 (Fig. S1), we give mAP at every IoU threshold from 0 to 1. The box of a slot model is cut from a 37×37 patch grid, so it is coarse. Even on a visible object, it reaches IoU above 0.5 on only 29.7% of the occlusions. Above 0.3 the metric measures the resolution of the grid and not the predicted position, so we average 0.1 to 0.3.
We also report a centre error that uses no threshold and no box size (response to Reviewer Ewxn, Question 2). There the centre predicted by iSEE stays within 0.34 to 0.59 object sizes of the ground truth under the static camera. RAM is at 0.31 to 0.55. This agrees with the mAP.
(c) Do most of the gains come from the walker, and is NoPE a precondition for the split?
We agree that position-free features are a precondition for the split. We have updated the paper to state this explicitly (Sec. 3.1 and Sec. 5).
We also agree that the walker, which follows RAM (Sec. 3.2), is responsible for the gain in permanence. However, the signal of occlusion, reappearance and identity comes from TEN and the two streams. They raise SAME from 32.2 to 82.0 before the walker is added (general response, G2). The gain is therefore possible only with the full combination, and not with the walker alone.
(d) Can slot pruning methods such as SOLV and DIAS be combined with TEN, and would pruning delete the slots TEN holds?
Thank you for the question and the references. Yes, they can be combined. SOLV and DIAS prune slots whose vectors are similar, by clustering on cosine distance. They do not prune on low attention. A held slot keeps the appearance of its hidden object, so it stays distinct from the other slots. TEN also marks every held slot (u in Eq. 6), so a pruning rule can skip them. We have added both references to the limitations/future work of the revised paper.