acceptodds
Under review as a conference paper at ICLR 2027

iSEE: Object Permanence Through Self-Supervision

Abstract

Object permanence, keeping track of an object’s identity and position while it is occluded, is central to video representations that track, predict and plan. Trackers that achieve it learn from boxes, track identities and visibility labels. On the other hand, self-supervised object-centric methods discover objects without labels: through slot attention, it represents a video as slots that bind to objects and follow them across frames. However, these slots are lost under occlusion, making the desired permanence impossible. Reasoning permanence is a hard problem because it requires to detect when an object becomes occluded, re-identify when object reappears, and keep the object’s hidden position continuous, using reapperance as the only learning cue. To address this, we propose iSEE, a novel framework that offers all three aforementioned requirements, without any labels whatsoever. We built iSEE using the following three proposed components: (i) Object evidence modelling: a slot’s attention, compared with its own past, reveals when its object is hidden. (ii) Appearance-position separation: two slot streams let the appearance be held for re-identification while the position keeps changing. (iii) Permanence from reappearance: a walker follows the hidden object’s position, trained only on where the object reappears. On LA-CATER static, iSEE returns a reappearing object to its own slot after 86 % of occlusions, against 32 % for SlotContrast (Manasyan et al., 2025), and localises it while hidden within 4.1 mAP of the label-trained SoTA RAM (Tokmakov et al., 2022). The two streams also allow downstream planning, with the position stream as the action of a world model.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
object permanence, object-centric learning, slot attention, self-supervised learning, occlusion, re-identification, video representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6V7MOpenReview ↗
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
  1. The diagnosis is clear and intuitive.
  2. It technique TEN is simple yet effective and parameter-free.
  3. Walker's training is elegantly self-supervised on appearance.
Critical weaknesses
  1. 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.
  2. 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.
  3. 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
AuthorsOpenReview ↗

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.

Thank you for the reply. Most of my concerns have been addressed. I keep my positive rating.

Reviewer_EwxnOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

iSEE extends slot-based video representation, enabling object identification and the estimation of object locations, even during periods of occlusion, without requiring ground-truth object maps for training. The algorithm decouples appearance from position, employs Temporal Evidence Normalization (TEN) to maintain or reset appearance states, and learns patch-based "walkers" grounded in re-emergence consistency. Its performance in object discovery, re-identification, and the localization of occluded objects was evaluated using the MOVi-C, LA-CATER, and KITTI datasets; the study also demonstrates its potential for latent planning. Experimental results show significant performance gains over SlotContrast, particularly in maintaining object identity within fixed-camera environments and re-identifying objects after prolonged occlusion.

Critical strengths
  1. This method provides a concrete explanation of how slots can maintain their identity even as their positions shift behind an occluder. In particular, the self-supervised walker proves highly effective. As shown in Table 4, the addition of the self-supervised walker improves the occluded-frame mAP from 30.1 to 80.9, and the SAME score from 82.0 to 86.3. These results demonstrate a significant impact not only on the detection of visible objects but also on the primary task of tracking objects while they are occluded.

  2. The evaluation was conducted by distinguishing between tracking occluded objects and re-identification, encompassing scenarios such as moving cameras, prolonged occlusion, and clear failure cases. On the static LA-CATER dataset, the SAME score improved from 32.2 to 86.3, while performance on KITTI re-identification rose from 29.0 to 48.4. These results can be interpreted more readily through the provided public scoring subset and the adjustments made to the baselines.

Critical weaknesses
  1. The role of the proposed representation decomposition is not sufficiently disentangled from the feature backbone. Table 4 introduces components sequentially: TEN, two-stream architecture, NoPE features, and the "walker" module. In the two-stream stage, the SAME metric drops from 77.6 to 71.0, but it recovers to 82.0 in the NoPE stage. Appendix B.1 clarifies that NoPE is not merely the removal of positional vectors during evaluation, but rather a separately fine-tuned feature extractor. Therefore, to properly assess the benefits of this representation design, it is essential to compare it against a "SlotContrast+TEN" baseline that utilizes the same NoPE features but lacks the decomposition. Table S6 provides support for disentangling contrastive loss from positional loss; the mAP for hidden layers is 30.1 when considering only appearance-based loss, compared to 13.0 when the full slot-based loss is applied. While this offers some evidence within the decomposed model, it cannot substitute for a direct comparison against a non-decomposed representation. Furthermore, Table S4 shows that SlotContrast+TEN already performs close to iSEE on the "static camera occluded objects" HOTA metric (81.3 vs. 82.1). While this does not negate the significant improvements in pedestrian detection, it suggests that the core architectural claims remain insufficiently validated.

  2. The evidence supporting robust, long-term persistence is more limited than what is suggested by the main results. The primary localization metrics rely on patch-grid readouts, resulting in IoU thresholds that average between 0.10 and 0.30 (Appendix B.4). Although the threshold curves and their upper bounds provide useful insights, these metrics conflate coarse localization estimates with mask size. In the static camera study, a subset of 195 events, deliberately selected for their high degree of motion, was used, while the KITTI re-identification test utilized only 39 events, employing target-centered crops (Appendices B.1–B.2). Although these configurations provide a reasonable diagnostic setting, they do not necessarily validate the test’s broader applicability and leave some uncertainty regarding both accuracy and robustness in real-world scenarios. An analysis of localization errors, broken down by mask coverage and source clip, would likely provide a more precise understanding of the extent of the improvements. Given these limitations, along with the absence of control over feature points, the core results appear promising, yet the actual gains remain marginal.

Questions for authors
  1. Table S6 supports loss decoupling, whereas Table 4 involves modifications to the backbone. Is there a SlotContrast+TEN control group that utilizes the same NoPE checkpoint and training schedule, while employing the SAME and hidden-location scores as they existed prior to the addition of the "walker"? (Weakness 1)

  2. The mAP metric in Appendix B.4 is calculated by combining center accuracy with mask coverage. To evaluate localization accuracy independently of bounding box size, could you report results that compare center error, normalized by object size, against occlusion duration, using standard event-based aggregation? (Weakness 2)

  3. Appendix B.4 applies a linear position map to visible frames from other clips. Please clarify which clips, target coordinates, and labels are used to fit this map, how they are distinguished from the evaluation set, and whether the baseline slot models used for comparison undergo the same calibration process. (Weakness 2)

  4. Table S2 features an augmented version of LA-CATER specifically focused on moving targets. Could you clarify the scope and robustness of these diagnostic improvements by separating the results for moving versus stationary events, and by reporting the variance in performance across different resamplings of the source clips? (Weakness 2)

  5. The KITTI study comprises 39 events extracted from 21 source videos (Appendix B.2). Could you provide confidence intervals for the improvements in re-identification, calculated by resampling the source videos while ensuring that correlated events remain grouped together? (Weakness 2)

  6. Appendix B.1 evaluates re-identification within a fixed crop area centered on the target. Please explain how these crop regions were determined, including whether annotations were used, and confirm whether all the methods under comparison were evaluated using the exact same crop regions. This will help clarify the implications of the results for full-scene tracking. (Weakness 2)

2 comments
AuthorsOpenReview ↗

Response to Reviewer Ewxn

Thank you for the careful reading and the detailed feedback.

Weakness #1 (HOTA): "Table S4 shows that SlotContrast+TEN already performs close to iSEE on the 'static camera occluded objects' HOTA metric (81.3 vs. 82.1). [...]"

We agree that the two methods are level on this row. This is expected, for two reasons.

First, HOTA is a whole-clip metric. Its association part (AssA) is averaged over every visible frame of the object, so an identity switch at one occlusion changes it by a small amount. HOTA is also the geometric mean of detection and association, so a gain in association enters at about half its size.

Second, under the static camera, Table S4 scores every occluded object, and 77% of them (172 of 224) do not move in the image while hidden. Holding the slot is enough for those. This is why the main paper limits such occlusions to 50 and keeps every occlusion through which the target moves (Sec. 4.1, Table S2).

To remove this bias, we computed HOTA on the same evaluation set as the main paper (static camera, 195 occlusions in 182 clips), with the same TrackEval protocol.

SlotContrastSlotContrast + TENiSEE
HOTA77.586.789.3
AssA65.082.089.6

All values ×100, higher is better. AssA is the part of HOTA that measures identity. On this set HOTA and SAME agree, and iSEE leads SlotContrast by 24.6 AssA. We have updated the paper appendix with this table, Appendix C.2 (Table S10).

Weakness #2 (metrics): "The evidence supporting robust, long-term persistence is more limited than what is suggested by the main results. [...]"

We address this weakness in Questions 2 to 6 below, with the centre error, the split into moving and stationary targets and the confidence intervals you asked for. We also ran a blinded user study on real-world video (general response, G3).

Question 1: Is there a SlotContrast + TEN baseline with the same NoPE features, scored without the walker?

We have added this ablation to Table 4 of the revised paper. Please refer to the general response (G2).

Question 2: Could you report the centre error, normalised by object size, against occlusion duration, with event-based aggregation?

We appreciate the suggestion. This metric measures the position of a hidden object without the size of the predicted box. We report it below on the occlusions of Table 1, with the occlusion lengths of Table 2. The error is the distance between the predicted centre and the centre of the ground-truth amodal box, divided by √(w·h) of that box (lower is better). It is averaged over the hidden frames of each occlusion, and then over the occlusions.

MethodStatic: 0.2–1.0 sStatic: 1.0–2.1 sStatic: 2.1–4.2 sStatic: ≥ 4.2 sMoving: 0.2–1.0 sMoving: 1.0–2.1 sMoving: 2.1–4.2 sMoving: ≥ 4.2 s
RAM0.360.310.550.490.450.480.650.92
RAM + refine0.350.160.390.280.430.390.540.76
frozen0.680.721.421.321.201.972.342.72
SlotContrast1.812.433.073.212.102.682.792.82
iSEE0.550.340.500.591.261.221.481.70

In short, under the static camera the error of iSEE does not grow with the occlusion length. It is 0.55 object sizes for occlusions under 1.0 s and 0.59 for occlusions of 4.2 s or longer, close to RAM at every length. The table therefore does not show a limit on long-term persistence as stated in Weakness 2.

Question 3: Which clips, target coordinates and labels are used to fit the position map, how are they separated from the evaluation set, and do the baseline slot models undergo the same calibration?

Thank you for the question. Yes, all slot baselines go through the same calibration as iSEE. Each of them uses the same estimator, the same frames and the same target, with its own slot as the input. RAM and Loci-Looped predict a box themselves, so no map is fitted for them.

The map is fitted on 100 clips of the LA-CATER training split, under the same camera as the evaluated set. All evaluated clips come from the test split, so no fitted clip is scored. The target is the centre of the ground-truth box of the object while it is visible, taken from the boxes released with LA-CATER. No hidden frame enters the fit.

Question 4: Could you separate the results for moving and stationary targets, and report the variance across resamplings of the source clips?

The table below splits the hidden-frame mAP of Table 1 by whether the target moves at least 10 px while it is hidden. The ± is the standard deviation over 10,000 resamplings of the source clips. We score every occlusion of all 1,371 test clips, so no occlusion is sampled (general response, G1).

MethodStatic camera, stationary target (606)Static camera, moving target (164)Moving camera, stationary target (243)Moving camera, moving target (402)
RAM98.8 ± 0.580.8 ± 1.891.3 ± 1.681.8 ± 1.2
RAM + refine98.2 ± 0.585.0 ± 1.988.9 ± 1.482.3 ± 1.2
frozen96.5 ± 0.643.1 ± 2.080.3 ± 2.126.2 ± 1.2
SlotContrast9.3 ± 0.76.3 ± 1.05.7 ± 1.01.6 ± 0.3
iSEE92.1 ± 1.078.4 ± 2.162.5 ± 2.744.1 ± 1.7

In short, under the static camera the gain of iSEE is on targets that move while hidden. There it is 35 points above frozen and close to RAM. On stationary targets every method except SlotContrast is near the ceiling.

Question 5: Could you provide confidence intervals for the re-identification gain on KITTI, resampling the source videos with correlated occlusions kept together?

We resample the 21 source videos with replacement 10,000 times. All occlusions of a video stay together, so correlated occlusions are never split. Both models are scored on the same resampled occlusions. The ± is the standard deviation over the resamples.

MethodSAME ↑ (%)95% confidence interval
SlotContrast29.0 ± 10.8[9.1, 51.4]
iSEE48.4 ± 8.8[33.3, 68.0]

In short, iSEE is ahead of SlotContrast in 85% of the resamples.

Question 6: How were the crop regions determined, were annotations used, and are all methods evaluated on the same crops?

We would like to clarify that all slot methods are evaluated on exactly the same crops. RAM is supervised with boxes, identities and visibility labels, and it receives the full frame. Each crop is picked once per occlusion from a fixed grid of 189×189 crops, using the annotated KITTI box centre on the last frame before the car hides and the first frame it is back. It stays fixed for the whole clip. We use a crop because a car that becomes fully occluded covers a median of only 4 patches on the full frame. In the crop it covers four times as many. The crop does not make the task easy, as SlotContrast reaches only 29.0 SAME on the same crops.

Reviewer_Ris7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SEE adds object permanence to SlotContrast, a self-supervised slot-attention video model, through three components:

  • Temporal evidence normalisation (TEN) detects when a slot's object is hidden and preserves its appearance until it reappears.
  • Separate appearance and position streams allow the model to retain an object's identity while updating its location.
  • A walker estimates the hidden object's position, learning from where it reappears rather than from labelled trajectories.

Experiments cover LA-CATER, MOVi-C, and KITTI (Tables 1–3), with component ablations (Table 4) and a planning demonstration (Sec. 4.4).

Critical strengths

(1) TEN (temporal evidence normalisation) works added to SlotContrast alone, it raises the share of occlusions after which the target returns to its own slot from 32.2% to 77.6% (Table 4). It is interesting as the idea could potentially can carry over to other slot models.

(2) The walker keeps track of where an object is while it is hidden, without using any labels (supported by table1). On frames where the target is fully occluded, adding the walker raises localisation accuracy from 30.1 to 80.9 mAP (Table 4). The improvement is noticeable in in Table 1 for occluded contained carried case with moving and static camera.

Critical weaknesses

None

Questions for authors

Requested clarification + evidence:

(1) How were the 195 and 388 scored occlusions chosen from the LA-CATER test set (Sec. 4.1)? Could you report the main comparison over all occlusions in the test set, with RAM evaluated as in its original paper?

(2) In Table 4, could you isolate the effect of the two-stream (appearance–position) split? On its own, adding the split lowers SAME from 77.6 to 71.0 and hidden-frame mAP from 19.6 to 6.9, and it only appears to help once NoPE features are added.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ris7

We thank the reviewer for the assessment and the constructive questions.

Question 1: How were the 195 and 388 scored occlusions chosen, and what are the results on the full test set?

We appreciate the suggestion. The selection rule is given in Appendix B.2 and Table S2. We have added one sentence to Sec. 4.1 of the revised paper that states which occlusions are scored. We have also added the comparison on the full test set, with RAM both raw and with its published post-processing. Please refer to the general response (G1).

Question 2: Can the effect of the two-stream split in Table 4 be isolated?

Please refer to the general response (G2). We trained SlotContrast + TEN on the same NoPE features without the split. Table 4 of the revised paper now adds the features before the split.

We are happy to answer any further questions.

Thank you for answering my questions. The clarifications and additional results address strengthen the manuscript. I will retain maintain my recommendation of weak acceptance, as the contributions are useful but the novelty is somewhat limited.

We thank the reviewer for the follow-up and the positive recommendation. We would like to briefly clarify the novelty.

iSEE is self-supervised, while RAM learns object permanence with labels (boxes, track identities and visibility labels). The novelty is a general principle: under occlusion, the appearance of an object must be held while its position keeps updating, so the two must be separate. TEN and the two-stream design implement this principle, and Table 4 verifies each step. Over the baseline SlotContrast, SAME rises from 32.2% to 86.3% and hidden-frame mAP from 6.6 to 80.9 on LA-CATER, which is significant.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers and the area chair for their time and feedback. Three points were raised by more than one reviewer. We answer them here as G1 to G3 and refer to them in the individual responses. All changes in the revised paper are marked in blue.

G1. Full LA-CATER test set (Reviewers Ris7, Ewxn, 6V7M). Table 1 of the paper is scored on 195 and 388 occlusions. As explained in Appendix B.2, these keep every occlusion through which the target moves and 50 in which it stays in place, because holding the last visible position already solves the latter. We agree with the reviewers that the full set should be reported as well, and do so below on every scorable occlusion of all 1,371 test clips per camera (770 static, 645 moving), with RAM both raw and with its published post-processing. However, as the table shows, the frozen baseline, which only holds the last visible position, already reaches 85.3 and 85.4 on occluded and contained frames under the static camera, and 59.8 on occluded frames under the moving camera. This is the reason for the selection described in Appendix B.2.

Table G1. Position of the object under occlusion on the full LA-CATER test set. mAP@[0.1,0.3] ↑ (×100) on the hidden frames of each label, on every scorable occlusion of all 1,371 test clips per camera (770 static, 645 moving). sup.: supervised with boxes, identities and visibility labels. frozen holds iSEE's slot at its last visible position. Bold: best row that reads the video alone.

Static: occludedStatic: containedStatic: carriedMoving: occludedMoving: containedMoving: carried
RAM + refine (sup.)92.997.388.784.689.182.3
RAM (sup.)91.995.582.285.387.779.7
frozen85.385.446.659.840.012.9
SlotContrast12.54.04.14.11.71.6
iSEE89.490.580.255.754.245.0

Under the static camera, iSEE is 2.0 to 5.0 points below raw RAM and 3.5 to 8.5 below RAM with post-processing, using no labels. Under the moving camera the gap to RAM remains large. We add this table to the appendix.

G2. Effect of the two-stream split with the features held fixed (Reviewers Ris7, Ewxn, 6V7M). In Table 4 of the paper the two stream split is added before the NoPE features, so its effect cannot be isolated with the features held fixed. We trained SlotContrast + TEN on the same NoPE features without the split, and now order Table 4 with the features first.

Table G2. Ablation on LA-CATER, static camera, with NoPE features added before the two-stream split. The first row is SlotContrast on NoPE features alone, without TEN. From the second row on, each row adds one component to the row above. SAME ↑ (%) and mAP ↑ (×100) on hidden frames and on all frames, as in Table 4.

MethodSAME ↑hidden mAP ↑mAP all ↑
SlotContrast + NoPE features30.11.12.6
SlotContrast32.26.614.6
+ TEN77.619.622.9
+ NoPE features79.23.43.7
+ two streams82.030.169.0
+ walker (iSEE)86.380.988.2

Removing position from the features already helps re-identification: SAME rises from 77.6 to 79.2. This gain needs TEN. Without it, SlotContrast on NoPE features reaches SAME 30.1, close to its 32.2 on DINOv2. NoPE features carry appearance only, so a reappearing object is matched to its slot by what it looks like and not by where it was. This is one of the findings behind our design, that the appearance features should be free of position. However, mAP suffers with NoPE features alone, because a single slot no longer carries the position. Our two-stream design improves re-identification further, to 82.0, with position as a separate stream. The walker cannot be applied trivially without the two streams. Table 4 of the revised paper follows this order, and Appendix C.7 reports further ablations, including the split on DINOv2 features and NoPE with position added back to a single slot.

G3. Real-world video: OVIS (Reviewers Ewxn, 6V7M). We trained SlotContrast and iSEE from scratch on OVIS for 100K steps each, with the same training videos, 8-frame training window, schedule and number of slots. We evaluate on a held-out set, the annotated training videos that contain a full occlusion, as the OVIS validation and test labels are not public. We report SAME (%, higher is better) by how far the object moved while hidden.

Methodunder 50 px50–150 px150 px or more
SlotContrast48.941.420.4
SlotContrast + TEN53.344.922.0
iSEE56.250.826.9

iSEE has the highest SAME in every range.

We ran a blinded user study to test object following and reappearance. 9 users each compared either iSEE and SlotContrast or iSEE and SlotContrast + TEN (only one pair per user) on 25 clips drawn at random from a pool of 95 KITTI and OVIS clips. Users preferred iSEE in 42.7% vs. SlotContrast in 15.7%, remaining scores in no preference. Similarly, users preferred iSEE in 41.8% vs. SlotContrast + TEN in 18.7%, remaining scores in no preference. The “no preference” is the suggested choice when both methods worked or both did not work.

Area_Chair_NmZEOpenReview ↗

Discussion with the authors

Dear reviewers,

the authors have provided responses for your reviews. Reviewer Ris7, thank you for letting the authors know that your concerns have been addressed. Can the others respond as well, please?

Thank you, Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.