acceptodds
Under review as a conference paper at ICLR 2027

Align2Refine: Diagnosing and Refining Video Representations through Spatial Latent Alignment

Abstract

Pre-trained video encoders can produce spurious tokens that fail to represent the visual content at their spatial locations. Prior work derives identification criteria from internal activations and token interactions, but extending these analyses to video is complicated by architectural diversity and complex spatiotemporal attention. We approach this challenge through task feedback: the spatial association between prediction errors and local feature fragmentation suggests that task responses can help locate spurious tokens. To obtain this feedback without downstream annotations, we formulate spatial latent alignment. We find that encoder representations capturing high-level semantics or low-level visual details can both serve as diagnostic references for this task. Building on these insights, we introduce Align2Refine, a framework for post-hoc refinement of video encoders on unlabeled videos. To make alignment feedback more discriminative for token identification, we use deliberately imbalanced learning to reinforce differences in fitting progress. The resulting token partition guides selective refinement using the original encoder's features as supervision. Across five heterogeneous video encoders, Align2Refine delivers substantial improvements in both spatial prediction and temporal correspondence under frozen-backbone evaluation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
video representation learning, post-hoc refinement, spatial latent alignment, dense representations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HSVnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Align2Refine, a post-hoc framework for improving pretrained video representations using unlabeled videos. A shared token-wise MLP predicts spatially corresponding features from a frozen reference encoder, and consistency-guided reweighting emphasizes tokens that are easier to align. A two-component Gaussian mixture calibrates a threshold for identifying regular and spurious tokens. Building on UniRefiner, the refinement stage trains student LoRA parameters to preserve regular teacher features and redirect spurious features into boundary registers, with the original encoder supplying the feature targets. Using only RGB frames from the YouTube-VOS training split for refinement, the paper evaluates five video encoders on semantic segmentation, monocular depth, video object segmentation, and point tracking under frozen-backbone evaluation. Reported gains include 4.4-21.5 ADE20K mIoU and 7.7-13.2 DAVIS-M J&F. Component ablations, alternative DINOv3/VAE references, and UPerNet evaluation provide additional evidence for the method.

Critical strengths

Broad and substantial empirical improvements (Tables 2-4 and 13). All five video encoders improve across the reported spatial and temporal tasks. V-JEPA 2 gains 11.0 ADE20K mIoU and 13.2 DAVIS-M J&F. Transfer to DAVIS and MOSE extends the evidence beyond the refinement dataset. UPerNet also improves by 2.3 mIoU

Critical weaknesses

W1. Diagnostic interpretation needs clearer qualification (Secs. 3-4.2; Table 1). The partition is useful empirically, but low consistency can also reflect ordinary sample difficulty or reference mismatch. Score bimodality is partly induced by the reweighting procedure. Table 1 supports an association with segmentation quality, although training-token counts, subset composition, and the shared evaluation population should be clarified. These issues limit the specificity of the spurious-token interpretation; they do not invalidate the mask's demonstrated practical usefulness. W2. Comparison with the closest refinement method is not fully controlled (Sec. 4.3; Appendix C, Table 10). The refinement framework is inherited from UniRefiner. Table 10 uses its published image-model results without fully establishing matched data and adaptation budgets, and no matched video comparison changes only the diagnostic criterion. This limits conclusions about superiority over the prior criterion. Nevertheless, the random-mask and reweighting ablations provide meaningful evidence for the proposed mechanism, so I regard this as a moderate comparative-evidence gap rather than a sufficient reason for rejection.

Questions for authors

Q1. In Table 1, are all three probes evaluated on the same complete validation set? Please specify training-token counts and subset composition. This would help distinguish diagnostic quality from differences in training-sample difficulty. Q2. Which refinement data, training budget, and evaluation settings are matched between the published UniRefiner results and your image-model results in Table 10? Is the original criterion transferable to the video refinement pipeline, and what evidence is available under otherwise matched settings? Q3. How stable are the learned thresholds and mask proportions across videos and crop-space recomputation? Please report available variability for the key Table 5 comparisons and the probe/refinement computational cost. These would strengthen confidence in the reported recipe.

5 comments

Thank you for the detailed suggestions. We have clarified Table 1 and added further experiments on diagnostic accuracy, method comparison, and stability.

Q1. Table 1 uses the same trained probe and validation set

(1) Table 1 evaluation protocol.

In Table 1, we train a single linear segmentation probe on all tokens from the entire ADE20K training set, without subset-specific sampling. The same probe is then evaluated on the complete validation set, with results reported separately for all, regular, and spurious token locations. Regular and spurious tokens account for 57.1% and 42.9% of the validation tokens, respectively.

(2) Token-level error validation.

To directly validate whether low-consistency tokens are more likely to produce spatial prediction errors, we evaluated regular and spurious tokens against ground-truth labels on Cityscapes and DAVIS-S using InternVideo-Next ViT-L.

Token-level accuracy (%) ↑RegularSpurious
Cityscapes91.6769.37
DAVIS-S (foreground)88.2156.45

Regular tokens achieve substantially higher prediction accuracy than spurious tokens on both benchmarks, showing that low-consistency tokens are much more likely to produce errors at their corresponding spatial locations. Importantly, the DAVIS-S results demonstrate that this relationship also holds in dynamic video scenes involving motion, occlusion, and appearance changes.

(3) Correlation between alignment consistency and dense prediction performance.

We also examined the relationship between raw alignment consistency and ADE20K segmentation accuracy (Appendix C, Figure 8). We grouped tokens by their cosine consistency scores and measured the prediction accuracy within each interval. The plot shows a clear, near-monotonic increase in accuracy as consistency rises, demonstrating that alignment consistency effectively reflects the quality of local token representations.

Q2. Align2Refine outperforms UniRefiner in matched video experiments

The original Table 10 reports published UniRefiner results for image encoders. We have now added a direct video-encoder comparison using matched refinement data, training epochs, LoRA rank, optimizer, input resolution, and downstream evaluation protocols.

Backbone / MethodADE20K mIoU ↑Cityscapes mIoU ↑VOC mIoU ↑NYUv2 RMSE ↓DAVIS-S J&F ↑
V-JEPA 2 (Original)25.453.465.20.54253.4
+ UniRefiner34.259.075.30.47761.9
+ Align2Refine36.462.277.90.45464.5
InternVideo-Next (Original)40.961.176.10.46360.7
+ UniRefiner42.262.978.00.44962.1
+ Align2Refine45.365.381.70.41166.6

Align2Refine performs better across all five metrics on both video encoders, showing the benefit of our alignment-based token identification under matched refinement settings.

Q3. Stable thresholds, crop-space masks, and lower computational cost

(1) Threshold and mask stability.

We measured threshold and mask stability on V-JEPA 2 ViT-g through repeated video sampling. For crop-space evaluation, we sampled 3 spatial crops per video and recomputed the masks using the same crop coordinates across frames.

StatisticMean ± Std
GMM threshold0.0614 ± 0.0037
Spurious-token ratio, full video (%)40.28 ± 1.36
Spurious-token ratio, spatial crops (%)41.49 ± 1.74

The threshold remains stable across sampled videos, while the spurious-token ratio stays similar between full videos (40.28%) and spatial crops (41.49%).

(2) Computational cost.

We also measured computational cost on the YouTube-VOS training split using 8 NVIDIA RTX 5090 GPUs under matched experimental settings.

BackboneMethodProbe time(min)Refinement time(min)Total time(min)
V-JEPA 2UniRefiner—77.477.4
Align2Refine12.946.759.6
InternVideo-NextUniRefiner—35.335.3
Align2Refine6.623.029.6

Including probe training, Align2Refine reduces total runtime by 23.0% on V-JEPA 2 and 16.1% on InternVideo-Next.

AuthorsOpenReview ↗

Hoping for a Discussion on Our Rebuttal for Paper 1130

Dear Reviewer,

I hope this message finds you well. We have carefully responded to the questions and concerns you previously raised, and we would greatly appreciate any further feedback you might be able to provide. Your insights are highly valuable to us, and we remain fully available to clarify or refine any remaining points.

As the discussion deadline approaches, we would be grateful for your response at your earliest convenience, as it would greatly help us move the discussion forward.

Thank you sincerely for your time and effort in reviewing our paper.

Best regards,

The Authors

Reviewer_c19mOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Align2Refine, a framework for identifying and refining spurious tokens in pretrained video encoders without annotations. Unlike prior methods that rely on internal feature and attention patterns, Align2Refine identifies spurious tokens based on how well they predict spatially aligned features from a reference encoder. The method uses deliberately imbalanced learning to better separate regular and spurious tokens, followed by selective refinement that preserves regular features while redirecting spurious features into boundary registers. Experiments on 5 video encoders show consistent improvements in semantic segmentation, depth estimation, video object segmentation, and zero-shot point tracking; ablations support the effectiveness of the proposed components.

Critical strengths
  • Effective method with strong and consistent experimental results. Align2Refine achieves substantial improvements across multiple downstream tasks (segmentation, depth estimation, point tracking). The qualitative results also paint a convincing picture of the method, showing more coherent features and segmentation masks with fewer holes and fragmented regions. Ablations further support the framework's effectiveness.
  • General applicability across different video encoders. The proposed diagnostic criterion does not rely on specific attention patterns, unlike past works such as UniRefiner [1], making it broadly applicable to different archtiectures. Effectiveness is demonstrated across 5 video encoders, with both semantic (DINOv3) and low-level visual features (SD3.5 VAE) as references.

[1] Qiu, Congpei, et al. "UniRefiner: Teaching Pre-trained ViTs to Self-Dispose Dross via Contrastive Register." arXiv preprint arXiv:2605.19622 (2026).

Critical weaknesses
  • Limited validation of spurious token identification. Low spatial alignment consistency does not necessarily indicate that a token is spatially corrupted/spurious, as it could also reflect difficulty predicting the reference representation. Table 1 show that the identified spurious tokens still achieve nontrivial segmentation performance. While the qualitative examples are convincing, a more systematic quantitative analysis of the relationship between alignment consistency and spatial prediction errors could help validate the proposed criterion and better clarify the role of imbalanced probe training.
  • Lack of computational efficiency analysis. The method requires training a separate alignment probe for each video encoder, followed by LoRA-based refinement with an external reference encoder. Although implementation details are provided, the paper does not report training times, GPU requirements, or computational costs. This makes it difficult to assess the approach's practicality, especially compared with UniRefiner [1], which emphasizes its lightweight post-hoc refinement.

[1] Qiu, Congpei, et al. "UniRefiner: Teaching Pre-trained ViTs to Self-Dispose Dross via Contrastive Register." arXiv preprint arXiv:2605.19622 (2026).

Questions for authors
  • Could the authors provide a quantitative analysis of how spatial alignment consistency relates to spatial prediction errors, and whether imbalanced probe training improves the identification of spurious tokens?
  • Could the authors report the training time and GPU requirements for both stages of Align2Refine and compare computational cost with UniRefiner?
5 comments

Thank you for the helpful suggestions. We have added direct token-level validation and a computational comparison with UniRefiner.

Q1. Alignment consistency reflects spatial prediction errors, and imbalanced training improves token identification

(1) Token-level error validation.

To directly validate whether low-consistency tokens are more likely to produce spatial prediction errors, we evaluated regular and spurious tokens against ground-truth labels on Cityscapes and DAVIS-S using InternVideo-Next ViT-L.

Token-level accuracy (%) ↑RegularSpurious
Cityscapes91.6769.37
DAVIS-S (foreground)88.2156.45

Regular tokens achieve substantially higher prediction accuracy than spurious tokens on both benchmarks, showing that low-consistency tokens are much more likely to produce errors at their corresponding spatial locations. Importantly, the DAVIS-S results demonstrate that this relationship also holds in dynamic video scenes involving motion, occlusion, and appearance changes.

(2) Correlation between alignment consistency and dense prediction performance.

We also examined the relationship between raw alignment consistency and ADE20K segmentation accuracy (Appendix C, Figure 8). We grouped tokens by their cosine consistency scores and measured the prediction accuracy within each interval. The plot shows a clear, near-monotonic increase in accuracy as consistency rises, demonstrating that alignment consistency effectively reflects the quality of local token representations.

(3) Effect of imbalanced probe training.

To validate that imbalanced probe training improves spurious-token identification, we compare the segmentation error rates of the identified spurious tokens on ADE20K, keeping all other settings unchanged. A higher error rate means that more of the identified tokens make incorrect segmentation predictions, indicating more accurate identification of spurious tokens.

Probe trainingSegmentation error rate of identified spurious tokens (%) ↑
w/o imbalanced training34.04
w/ imbalanced training38.53

With imbalanced training, the identified spurious tokens have a higher error rate, showing that our training strategy more effectively identifies spurious tokens.

Q2. Practical and efficient refinement of video encoders

(1) Training efficiency.

Align2Refine enables efficient refinement of large-scale video encoders. Using 8 NVIDIA RTX 5090 GPUs, we can refine V-JEPA 2 ViT-g within one hour, including alignment-probe training. We further compare its computational cost with UniRefiner under matched settings on the YouTube-VOS training split used in our main experiments.

BackboneMethodProbe time(min)Refinement time(min)Total time(min)
V-JEPA 2UniRefiner—77.477.4
Align2Refine12.946.759.6
InternVideo-NextUniRefiner—35.335.3
Align2Refine6.623.029.6

Including probe training, Align2Refine reduces total runtime by 23.0% on V-JEPA 2 and 16.1% on InternVideo-Next.

(2) Why Align2Refine is faster.

The main difference lies in how the methods identify spurious tokens. UniRefiner concatenates unrelated images, requiring larger inputs, and uses more complex detection procedures, including FP/GP matching and multilayer attention-hijackee (AH) statistics. These operations are particularly expensive for video encoders.

Align2Refine uses learned alignment with reference features to identify tokens, providing a cheaper screening signal and reducing the cost of generating supervision during refinement.

(3) Downstream performance.

We also compared downstream performance under the same matched settings.

Backbone / MethodADE20K mIoU ↑Cityscapes mIoU ↑VOC mIoU ↑NYUv2 RMSE ↓DAVIS-S J&F ↑
V-JEPA 2 (Original)25.453.465.20.54253.4
+ UniRefiner34.259.075.30.47761.9
+ Align2Refine36.462.277.90.45464.5
InternVideo-Next (Original)40.961.176.10.46360.7
+ UniRefiner42.262.978.00.44962.1
+ Align2Refine45.365.381.70.41166.6

Align2Refine achieves better downstream performance at lower total training cost in these matched video-encoder experiments.

AuthorsOpenReview ↗

Hoping for a Discussion on Our Rebuttal for Paper 1130

Dear Reviewer,

I hope this message finds you well. We have carefully responded to the questions and concerns you previously raised, and we would greatly appreciate any further feedback you might be able to provide. Your insights are highly valuable to us, and we remain fully available to clarify or refine any remaining points.

As the discussion deadline approaches, we would be grateful for your response at your earliest convenience, as it would greatly help us move the discussion forward.

Thank you sincerely for your time and effort in reviewing our paper.

Best regards,

The Authors

Reviewer_exdnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies spatially misleading local tokens in pre-trained video encoders and proposes Align2Refine, a post-hoc representation refinement framework trained on unlabeled videos. Motivated by the observed spatial overlap between segmentation errors and fragmented features, it formulates spatial latent alignment to diagnose token quality without downstream annotations. A frozen video encoder and a reference encoder (DINOv3 or SD3.5 VAE) provide spatially matched representations. A shared MLP is trained with a consistency-reweighted InfoNCE objective, and a two-component Gaussian mixture model separates tokens into regular and spurious groups. During refinement, a frozen teacher supervises a LoRA-adapted student: regular tokens are preserved at corresponding locations, while spurious teacher features supervise temporal boundary registers. Across five heterogeneous video encoders refined on RGB videos from YouTube-VOS, Align2Refine improves frozen-backbone semantic segmentation, monocular depth estimation, video object segmentation, and zero-shot point tracking. Component ablations and UPerNet experiments offer additional evidence of the method’s effectiveness.

Critical strengths
  1. The paper presents an architecture-agnostic way to diagnose local video representations using spatial latent alignment rather than encoder-specific attention or activation heuristics. The use of both a semantic reference (DINOv3) and a low-level reference (SD3.5 VAE) provides useful evidence that the diagnostic task is not tied to one type of reference representation.
  2. The empirical evaluation is broad and largely controlled: all five video encoders improve on spatial and temporal tasks under frozen-backbone protocols, while Tables 2–5 provide multi-dataset results and component ablations. The additional UPerNet evaluation shows that gains are not restricted to a linear segmentation probe.
Critical weaknesses
  1. Insufficient validation of the spurious-token criterion: In Sec. 4.2, low alignment consistency is directly interpreted as evidence of a spurious token. However, disagreement with DINOv3 or SD3.5 VAE may also reflect useful video-specific information, motion, occlusion, or reference-model limitations. Table 1 only shows that the selected low-consistency subset has weaker segmentation probing performance on one backbone; it does not independently establish token-level spuriousness. The authors should validate the detected locations against localized errors or controlled token artifacts, particularly for dynamic scenes.
  2. Limited attribution of gains beyond the inherited refinement framework: Sec. 4.3 explicitly adopts UniRefiner’s preservation-and-absorption framework. Yet the UniRefiner comparison in Appendix C is limited to image encoders, with UniRefiner scores taken from the earlier paper, rather than a matched video-encoder experiment. More importantly, Table 5 shows that a random mask already improves ADE20K mIoU from 25.4 to 33.4, compared with 36.4 for the full method. Thus, a substantial share of the gain may come from generic refinement. Matched comparisons with UniRefiner and simpler LoRA/distillation or register baselines on video encoders are needed.
  3. Limited validation across refinement data distributions In Section 5.1 and Appendix A.1, all five video encoders are refined exclusively on the YouTube-VOS training split. Although the authors demonstrate improvements across multiple downstream benchmarks, these experiments do not establish whether Align2Refine remains effective when the unlabeled refinement data come from different distributions. Since the framework is intended for general post-hoc refinement of video representations, the authors should evaluate its performance using alternative unlabeled video datasets to verify that the observed improvements are not dependent on the characteristics of YouTube-VOS.
Questions for authors

See the analysis of weaknesses above.

3 comments

Thank you for the constructive feedback. We have added further experiments to address the three concerns.

Q1. Alignment consistency is closely related to spatial prediction errors

(1) Token-level error validation.

To directly validate whether low-consistency tokens are more likely to produce spatial prediction errors, we evaluated regular and spurious tokens against ground-truth labels on Cityscapes and DAVIS-S using InternVideo-Next ViT-L.

Token-level accuracy (%) ↑RegularSpurious
Cityscapes91.6769.37
DAVIS-S (foreground)88.2156.45

Regular tokens achieve substantially higher prediction accuracy than spurious tokens on both benchmarks, showing that low-consistency tokens are much more likely to produce errors at their corresponding spatial locations. Importantly, the DAVIS-S results demonstrate that this relationship also holds in dynamic video scenes involving motion, occlusion, and appearance changes.

(2) Correlation between alignment consistency and dense prediction performance.

We also examined the relationship between raw alignment consistency and ADE20K segmentation accuracy (Appendix C, Figure 8). We grouped tokens by their cosine consistency scores and measured the prediction accuracy within each interval. The plot shows a clear, near-monotonic increase in accuracy as consistency rises, demonstrating that alignment consistency effectively reflects the quality of local token representations.

Q2. Align2Refine outperforms UniRefiner on video encoders under matched settings

(1) Matched comparison with UniRefiner.

We compared both methods on two video encoders, using the same refinement data, training epochs, LoRA rank, optimizer, input resolution, and downstream evaluation protocol.

Backbone / MethodADE20K mIoU ↑Cityscapes mIoU ↑VOC mIoU ↑NYUv2 RMSE ↓DAVIS-S J&F ↑
V-JEPA 2 (Original)25.453.465.20.54253.4
+ UniRefiner34.259.075.30.47761.9
+ Align2Refine36.462.277.90.45464.5
InternVideo-Next (Original)40.961.176.10.46360.7
+ UniRefiner42.262.978.00.44962.1
+ Align2Refine45.365.381.70.41166.6

Align2Refine achieves the best results across all five metrics on both video encoders.

(2) All-token distillation baseline.

We also compared it with all-token distillation on V-JEPA 2 ViT-g. This baseline aligns all tokens without registers or absorption loss.

MethodADE20K mIoU ↑NYUv2 RMSE ↓DAVIS-S J&F ↑
Original25.40.54253.4
All-token distillation32.30.49859.1
UniRefiner34.20.47761.9
Align2Refine36.40.45464.5

Align2Refine consistently outperforms both generic feature distillation and UniRefiner. These results highlight the benefit of our alignment-based token identification beyond existing refinement approaches.

Q3. Improvements are consistent across refinement datasets

We repeated refinement using the unlabeled MOSE training split instead of YouTube-VOS, without using segmentation annotations.

BackboneRefinement dataADE20K mIoU ↑DAVIS-S J&F ↑
V-JEPA 2YouTube-VOS36.464.5
MOSE36.264.1
InternVideo-NextYouTube-VOS45.366.6
MOSE45.166.3

Performance remains nearly unchanged across the two refinement datasets, showing that the improvements do not depend on YouTube-VOS specifically.

The revised manuscript has substantially addressed the three major concerns raised in my initial review.

First, the token-level accuracy results on Cityscapes and DAVIS-S (Table 12), together with the consistency–accuracy analysis in Figure 8, provide stronger empirical evidence that low alignment consistency is associated with local prediction errors. These results adequately support alignment consistency as a practical diagnostic criterion, although they do not necessarily establish that all low-consistency tokens are intrinsically spurious.

Second, the matched comparisons with UniRefiner on two video encoders (Table 14) and the additional all-token distillation baseline (Table 15) directly address the previous concerns regarding baseline comparison and performance attribution. The consistent improvements across multiple evaluation metrics demonstrate the additional benefit of the proposed alignment-based diagnostic strategy beyond existing refinement approaches.

Third, the experiments using MOSE as an alternative refinement dataset (Table 7) show comparable performance to YouTube-VOS, sufficiently addressing the concern about dependence on a particular refinement dataset.

The remaining limitations are relatively minor. In particular, the distinction between alignment consistency as an empirical indicator and the underlying causes of spurious representations should be maintained. The token-level accuracy evaluation in Appendix C could also be described more explicitly to improve reproducibility.

Overall, I appreciate the authors' efforts to strengthen the empirical validation. The additional experiments have substantially resolved my original concerns and improved the paper's technical credibility. I therefore update my recommendation from Weak Rejection to Weak Acceptance.

We sincerely thank the reviewer for the positive reassessment and for recognizing that our additional experiments address the main concerns.

We agree with the remaining suggestions and have made two clarifications in the revised manuscript:

  1. Sec. 4.2: We clarify that low alignment consistency is an empirical indicator of spatial prediction reliability, without implying that every identified token is intrinsically spurious.
  2. Appendix C: We provide a more explicit description of the token-level evaluation procedure to improve reproducibility.

We appreciate the reviewer's constructive feedback, which has helped us strengthen the paper.

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged that the reviewers recognize the following strengths of our work:

  • General applicability: An architecture-agnostic diagnostic approach that works across diverse video encoders and reference representations.
  • Strong performance: Consistent improvements across five video encoders on spatial and temporal tasks.
  • Comprehensive evaluation: Extensive experiments, convincing visualizations, and ablation studies supporting the framework's effectiveness.

We have addressed the reviewers' main concerns through additional experiments and analyses. Changes included in the revised manuscript are highlighted in blue:

  • Alignment consistency and token quality (exdn, c19m, HSVn): Regular tokens achieve substantially higher prediction accuracy than spurious tokens on Cityscapes and DAVIS-S. A new plot (Appendix C, Figure 8) shows that segmentation accuracy increases nearly monotonically with alignment consistency, demonstrating that consistency effectively reflects local token representation quality. Additional experiments show that imbalanced probe training improves spurious-token identification.
  • Matched comparisons(exdn, HSVn): Under matched video-refinement settings, Align2Refine outperforms UniRefiner across all five metrics on video encoders and also surpasses all-token distillation (Appendix D).
  • Generalization and stability (exdn, HSVn): Refinement on MOSE achieves nearly identical results to YouTube-VOS. The learned thresholds and token-selection proportions also remain stable across videos and spatial crops (Appendix A).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.