acceptodds
Under review as a conference paper at ICLR 2027

ReMoM: Relation-aware Motion Matching for Motion-Guided Few-Shot Video Object Segmentation

Abstract

Motion-guided few-shot video object segmentation (M-FSVOS) aims to segment objects in a query video that exhibit motion patterns defined by a few support videos. A key challenge of M-FSVOS lies in support-defined motion correspondence, i.e., reliably matching motion across support and query videos despite variations in appearance and execution style. Although state-of-the-art video object segmentation models such as SAM 3 excel at object discovery, segmentation, and tracking, their object representations rely heavily on appearance cues and thus fall short in the M-FSVOS task. To tackle this challenge, we introduce a novel model named Relation-aware Motion Matching (ReMoM), which formulates M-FSVOS as a motion-centric masklet matching problem. Under this formulation, ReMoM performs support-conditioned motion matching through explicit track-to-way relation modeling over high-quality masklets extracted by SAM 3. Additionally, it strengthens the temporal evidence of the learned motion relations through counterfactual temporal enhancement and incorporates interaction-relevant context from surrounding objects to resolve interaction-dependent ambiguities. Extensive experiments on the MOVE and A2D-Motion benchmarks demonstrate the effectiveness of ReMoM, which significantly outperforms state-of-the-art models by 17.5% on MOVE.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Object Segmentation, Motion-Guided Few-Shot Video Object Segmentation

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

This paper proposes ReMoM for motion-guided few-shot video object segmentation. It uses SAM 3 to generate object tracks and introduces three modules for motion matching, temporal enhancement, and interaction-aware context modeling.

Critical strengths
  1. The paper is well organized, and the motivation is clear.
  2. The masklet matching framework is interesting and makes good use of SAM3.
  3. The experimental results are impressive, with consistent improvements on MOVE and A2D-Motion.
Critical weaknesses
  1. The three modules are based on relatively common techniques, such as feature difference, temporal shuffling, and cross-attention. Their unique contributions could be explained more clearly.
  2. The method relies heavily on SAM3 for generating candidate masklets, but the impact of masklet quality on the final performance is not well discussed.
  3. The experiments are generally sufficient, but some additional analysis of the model's motion understanding ability would make the results more convincing.
Questions for authors
  1. Why does the interaction-aware module only use appearance features from surrounding objects instead of also considering their motion features?
  2. How sensitive is the method to the quality of the masklets generated by SAM 3?
3 comments
AuthorsOpenReview ↗

Response to Reviewer fUEU (1/1).

W1. Unique contributions of the three proposed modules.

Answer to W1: We would like to clarify that our central contributions are the motion-centric masklet matching formulation and three complementary modules for support-defined motion discrimination, providing a framework for equipping SAM 3 with motion understanding capabilities tailored to M-FSVOS. The feature difference, temporal shuffling, and cross-attention just serve as building blocks within our track-to-way relation encoder, counterfactual temporal enhancement, and interaction-aware contextualization, respectively. In what follows, we clarify the contributions of the three proposed modules, each of which is designed to address a unique challenge.

  • Track-to-way relation encoder. Since motion categories are defined by support exemplars, our encoder explicitly models the relation between query tracks and support motions, capturing fine-grained motion agreement and discrepancy for support-specific discrimination.

  • Counterfactual temporal enhancement. This module improves motion understanding by enhancing sensitivity to temporal order. It contrasts the factual query with an appearance-preserving shuffled view, and uses their relation difference to strengthen temporally sensitive motion evidence. By capturing order-sensitive information, this module strengthens motion modeling for support-conditioned matching.

  • Interaction-aware contextualization. In M-FSVOS, many interaction-dependent motions require context from surrounding objects for discrimination. This module selectively aggregates information from surrounding objects, injecting interaction-relevant context that the target track alone does not contain, and therefore helps resolve the ambiguities in interaction-dependent motions.

Experiments further demonstrate the effectiveness of our formulation and module designs, with ReMoM consistently outperforming SAM 3-based baselines and previous state-of-the-art methods on MOVE and A2D-Motion.

W2 / Q2. Sensitivity to SAM 3 masklet quality.

Answer to W2 / Q2: Thanks for pointing out the insufficient discussion of masklet quality. We conduct two experiments: independently dropping each candidate masklet with probability pp to simulate missed tracks, and randomly masking a fraction rr of temporal object tokens to simulate missing observations associated with fragmented tracks.

Masklet qualityProbability / fractionJ&F
Original masklet—68.2
Missed objectsp=0.0565.5
Missed objectsp=0.1063.0
Missed objectsp=0.1560.5
Fragmented tracksr=0.1067.2
Fragmented tracksr=0.2065.8
Fragmented tracksr=0.3064.5
DMA-VSwin-T (SOTA)—51.1

ReMoM retains 60.5 J&F with 15% missed objects and 64.5 with 30% fragmented tracks. These results demonstrate that ReMoM still outperforms the previous state of the art, DMA–VSwin-T (51.1 J&F), supporting its effectiveness and robustness with low masklet quality.

W3. Additional analysis of motion understanding ability.

Answer to W3: Thank you for this suggestion. We conduct additional motion analysis experiments with module ablations to investigate whether the model captures motion cues. Specifically, we compare factual queries with static queries that remove motion variation while retaining appearance cues.

ModelFactualStaticΔ\Delta
Relation64.558.1-6.4
Relation + context66.760.8-5.9
ReMoM68.259.5-8.7

All three configurations perform worse with static queries, supporting that they capture motion cues beyond static appearance.

Q1. Why use only appearance features as contextual keys and values?

Answer to Q1: We provide an ablation on features used as cross-attention keys and values to clarify this design, and the results are shown below. Using only motion features decreases J&F by 3.0 points. These results support our choice of appearance features as contextual keys and values.

Keys and valuesJ&FΔ\Delta
Appearance68.2—
Motion65.2-3.0

We conjecture that this is because the model should focus on the target’s own motion, while surrounding-object appearance provides complementary cues for distinguishing interaction types.

Reviewer_fUEUOpenReview ↗

Response to Author

Thank you for your careful responses and we tend to retain the rating

Dear Reviewer fUEU,

We are encouraged by your recognition of our response. We sincerely appreciate your dedicated time and effort in offering invaluable feedback.

Best

Reviewer_snFPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReMoM for motion-guided few-shot video object segmentation (M-FSVOS). The method reformulates the task as motion-centric matching over object tracks extracted by a frozen SAM 3 model. It introduces track-to-way relation modeling, counterfactual temporal enhancement, and interaction-aware contextualization to improve motion discrimination across support and query videos. Experiments on MOVE and A2D-Motion show substantial improvements over existing methods.

Critical strengths
  • Effective task reformulation. Formulating M-FSVOS as object-track matching is a reasonable way to leverage the strong segmentation and tracking capabilities of SAM 3 while focusing on motion correspondence. This separates object segmentation from motion-based target selection.

  • Well-motivated motion matching. The proposed components address different aspects of motion discrimination, including support-query correspondence, temporal ordering, and interactions with surrounding objects. The design is coherent, and the ablation results support the contributions of these components.

  • Strong experimental results. ReMoM consistently outperforms previous methods on MOVE under both overlapping and non-overlapping splits. The comparisons with SAM 3-based baselines are particularly convincing, and the additional results on A2D-Motion provide further evidence of generalization to unseen motions.

Critical weaknesses
  • Limited novelty of the individual components. The relation encoder, temporal shuffling, gated residual enhancement, and cross-attention largely build on established techniques. While their combination is effective, the paper could better clarify what is technically distinctive about the proposed motion matching mechanism beyond adapting these components to SAM 3 object tracks.

  • Insufficient analysis of motion versus appearance cues. The interaction-aware module aggregates appearance features from surrounding objects rather than explicitly modeling their motion relationships. It is therefore unclear whether its improvement comes from understanding interactions or simply recognizing contextual objects. Moreover, the static-query experiment still achieves 61.1 J&F compared with 69.0 for the full model, suggesting that appearance cues remain influential. More controlled experiments that separate motion and appearance information would strengthen the conclusions.

  • Limited analysis of candidate quality and efficiency. ReMoM relies on masklets generated by SAM 3, but the effects of missed objects, fragmented tracks, and imperfect masks are not systematically evaluated. The relatively low null-way rejection accuracy (23.1%) also deserves attention. Reporting candidate recall, end-to-end inference time, and computational overhead would help assess the robustness and practicality of the framework.

Questions for authors

Questions

  1. Could the authors further clarify the novelty of the track-to-way relation encoder compared with conventional feature matching or relation learning methods? Would a simpler matching strategy achieve comparable results using the same SAM 3 features?

  2. How much does interaction-aware contextualization rely on surrounding object appearance rather than their motion relationships? Could the authors provide a more controlled analysis to distinguish these two sources of information?

  3. How sensitive is ReMoM to the quality of SAM 3-generated masklets? In particular, how do missed objects or fragmented tracks affect performance? Could the authors also report the end-to-end inference time and computational overhead?

  4. How would ReMoM perform if SAM 3 were replaced with SAM 2? Since the framework relies heavily on SAM 3-generated masklets and object tokens, a comparison using SAM 2 with corresponding object representations would help clarify whether the improvements mainly come from the proposed motion matching mechanism or the stronger SAM 3 backbone.

5 comments
AuthorsOpenReview ↗

Response to Reviewer snFP (1/3).

W1. Technical Distinctiveness of the Proposed Components.

Answer to W1: We would like to clarify that our primary novelty lies in reformulating M-FSVOS as a motion-centric masklet matching problem, which fundamentally differs from the formulation adopted by previous methods. Building upon this novel formulation, we propose ReMoM, an effective model that equips SAM 3 with motion understanding capabilities tailored to M-FSVOS. The components in ReMoM are specifically designed to address the unique challenges of M-FSVOS. Below, we provide point-by-point responses to the designs mentioned by the reviewer:

  • Track-to-way relation encoder. We specifically design this module for support-defined motion discrimination. In M-FSVOS, motion categories are specified by support motion exemplars, so recognizing a query motion requires explicitly comparing it against the demonstrated support motions. To this end, our relation encoder explicitly models the relationships between query tracks and support motion ways, enabling support-specific motion discrimination. Unlike conventional prototype similarity matching, our approach learns support-conditioned relational representations rather than relying on a simple similarity score.

  • Temporal shuffling and gated residual enhancement. We would like to clarify that we do not claim temporal shuffling or gated residual connections themselves as our technical contribution. They are just building blocks of our counterfactual temporal enhancement. Specifically, our counterfactual temporal enhancement strengthens temporal evidence for motion understanding by contrasting the motion relations of the factual query and an appearance-preserving shuffled view, and leveraging their difference to enhance temporal sensitivity. Experiments demonstrate that our counterfactual temporal enhancement module improves J&F from 67.9 to 69.0, supporting its effectiveness.

  • Cross-attention. We would like to clarify that we do not claim cross-attention itself as a technical contribution, as it is just a building block of our interaction-aware contextualization. Specifically, our interaction-aware contextualization module is designed for the unique challenge of M-FSVOS, that many interaction-dependent motions require context from surrounding objects for discrimination. Our module selectively aggregates information from surrounding objects, injecting interaction-relevant context that the target track alone does not contain, and therefore helps resolve the ambiguities in interaction-dependent motions.

Experiments further demonstrate the effectiveness of our formulation and module designs, with ReMoM consistently outperforming SAM 3-based baselines and previous state-of-the-art methods on MOVE and A2D-Motion.

W2. / Q2 More analysis for contextualization and importance of motion cues.

Answer to W2 / Q2: Following the reviewer’s comments, we conduct an attention-query ablation, and the results below support that contextualization uses motion information in addition to surrounding-object appearance. Specifically, we remove the motion inputs from the contextual attention query while retaining its appearance inputs and the surrounding-object features used as keys and values.

Motion inputs in the attention queryJ&FΔ\Delta
Query and support motion: (mqm^q, msm^s)68.2—
No motion inputs67.2-1.0

Removing both query and support motion reduces J&F by 1.0 point. These results demonstrate that the interaction-aware module leverages the interaction motion rather than simply recognizing the surrounding objects.

Second, we would like to emphasize that motion understanding ability is more important and essential for M-FSVOS, although we acknowledge that appearance cues also contribute some performance . To further investigate motion understanding beyond appearance cues, we evaluate the three query type of factual, shuffled and static across module ablations to verify each module’s motion understanding ability.

ModelFactualShuffleStatic
Relation64.563.9 (−0.6)58.1 (−6.4)
Relation + context66.766.0 (−0.7)60.8 (−5.9)
ReMoM68.264.4 (−3.8)59.5 (−8.7)

These additional analyses show that disrupting or removing motion cues consistently reduces performance across all model configurations, supporting the importance of motion cues for all of our modules.

AuthorsOpenReview ↗

Response to Reviewer snFP (2/3).

W3. / Q3. The effect of mask quality, null way rejection report, candidate recall report, and computational cost measurement.

Answer to W3 / Q3: We provide additional analysis on candidate quality, comparisons of null-way rejection accuracy, candidate recall report, and measurements of inference latency and computational overhead.

Candidate quality. We conduct three experiments to evaluate robustness to SAM 3 masklets quality: dropping candidates to simulate missed objects, masking temporal tokens to simulate fragmented tracks, and perturbing masks to simulate imperfect masks.

Candidate qualityProbabilityJ&F
Original masklet—68.2
Missed objects0.0565.5
Missed objects0.1063.0
Missed objects0.1560.5
DMA-VSwin-T (SOTA)—51.1
Candidate qualityRatioJ&F
Original masklet—68.2
Fragmented tracks0.1067.2
Fragmented tracks0.2065.8
Fragmented tracks0.3064.5
DMA-VSwin-T (SOTA)—51.1
Candidate qualityJ&F
Original masklet68.2
Boundary noise (10%)67.7
Eroded masks65.2

These results show that degraded masklet quality causes some performance loss. However, ReMoM still outperforms the previous state of the art method DMA–VSwin-T (51.1 J&F), supporting ReMoM’s effectiveness and robustness even with imperfect SAM 3 masklets.

Null-way rejection. We provide additional N-Acc comparisons with baseline methods under the same MOVE OS 2-way-1-shot setting.

MethodN-Acc (%)
DMA–VSwin-T (SOTA)21.2
SAM 3 + TimeAttn7.3
SAM 3 + SOLA6.0
ReMoM23.1

All methods achieve relatively low N-Acc, reflecting the difficulty of null-way rejection, because the model must reject any motion outside the support-defined categories, without explicit examples covering these out-of-support motions. Despite this difficulty, ReMoM achieves the highest N-Acc of 23.1%, surpassing the previous state of the art (21.2%) and improving over SAM 3 + TimeAttn and SAM 3 + SOLA by 15.8% and 17.1%, respectively. The results demonstrate that ReMoM has relatively strong null-way rejection performance.

Candidate recall. We evaluate candidate recall on the test set at IoU thresholds of 0.5 and 0.75, obtaining 95.6% and 93.1%, respectively. These high recall rates indicate that SAM 3 excels at object discovery and tracking (though weak in motion understanding).

Computational cost. We benchmark ReMoM and SAM 3 + TimeAttn baseline on the same 20 test episodes using a single RTX 5090, and the results are shown below.

ModelForward latency (s/ episode)FLOPs (TFLOPs/episode)J&F
SAM 3 + TimeAttn97.035611.6061.4
ReMoM97.045611.6169.6

Compared with SAM 3 + TimeAttn, ReMoM introduces negligible additional computational cost while improving J&F by 8.2 points. This substantial performance gain with minimal overhead demonstrates the effectiveness of our design.

AuthorsOpenReview ↗

Response to Reviewer snFP (3/3).

Q1. Comparison with simple matching.

Answer to Q1: Our method significantly differs from conventional feature matching or relation learning (prototype learning). Our track-to-way relation encoder is explicitly designed to capture fine-grained motion agreement and discrepancy, providing a discriminative relational representation rather than a similarity score. This takes inspiration from that the motion categories are defined by support exemplars, so recognizing a query motion requires explicitly comparing it against the demonstrated support motions. We compare our relation encoder version with SAM 3 Raw, which uses cosine similarity between temporally averaged tokens, and SAM 3 + TimeAttn, a feature-matching baseline that applies temporal attention followed by prototype similarity.

MethodJ&F
SAM 3 Raw30.6
SAM 3 + TimeAttn60.0
Relation encoder66.3

Our relation encoder improves over these baselines by 35.7 and 6.3 points, respectively, supporting the effectiveness of our track-to-way relation encoder.

Q4. Replacing SAM 3 with SAM 2.

Answer to Q4: We clarify that our improvements mainly come from the proposed motion matching module, whose effectiveness extends beyond SAM 3. To verify this, we replace SAM 3 with SAM 2 for masklet generation and object token extraction, and evaluate on MOVE under the OS 2-way-1-shot setting.

MethodJ&F
DMA–VSwin-T51.1
SAM 2 + TimeAttn46.7
SAM 2 + ReMoM60.0

SAM 2 + ReMoM outperforms SAM 2 + TimeAttn by 13.3 points, demonstrating the benefit of our motion-matching module under the same backbone. By replacing SAM 3 with SAM 2, ReMoM still substantially surpasses the previous state of the art, DMA–VSwin-T, by 8.9 points. These results support our design’s effectiveness in enabling support-defined motion understanding across different SAM backbones.

Reviewer_t5qFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This manuscript argues that being able to segmentation video object based on motion pattern is an important step towards effective few-show video object segmentation, whereas existing methods leverage both object appearence and temporal dynamic. To achieve this, a motion centric matching method built upon a powerful foundation model (i.e, SAM30 is proposed to validate an adapted motion-oriented benchmark (from A2D dataset).

Three interesting designs are proposesd: 1) an additional motion dynamic modelling implemented as temporal-wise feature difference, 2) temporally dis-ordered counterfactual motion feature set, and 3) the interaction modelling between objects. Arguably, all these operation could be modelled implicitly with a transformer model if the information source are inputed together. And later on, in the experiment section, these prior-driven designs are proven useful.

The results on MOVE dataset, which contains the most motion category, is quite notable - surpassing the previous Sota.

Critical strengths
  • Appreciated the cleanness in writing (i.e., math), experimental hyper-parameters sharing, and result presentation.

  • The solution, which is mostly prior-knowledge driven design, is intuitive.

  • The results are notable and seem to be a signficant milestone on this research direction (assuming there exists few/no critical bug in evaluation and the model training is reproducible).

Critical weaknesses
  • No critical weakness from this reviewer.

  • Minor 0: Could you pls provide some justification on the SAM3 model was not pre-trained on the test set of MOVE/A2D dataset, to be more rigorous? (Understanding that this ablation study won't challenge the main conclusion - the proposed method is effective).

  • Minor 1: Lack of deeper understanding of what type of motion could be captured by the proposed method. More importantly, the contrast between motion capture by SAM3 and proposed method.

  • Minor 2: In the era of foundation model, the zero-shot performance using SAM3 exposes the foundtation model is still not motion-centric enough, at least not good enough for this task. It might worth provide reflection / future direction on what could further improve the foundation model training recipe, which could be stronger take-way message to the community than bold number over benchmarks.

Questions for authors
  • How is \textbf{W} (in the W×K support masklets) decided?

  • What is the intuition of chosing TCN?

  • What do you think about the 'termporal order shuffling' assumption?

    • For example, how about \textit{Jumping jacks} which is bit more symmetric?
  • Could you be more clear about the definition of the research topic "motion matching", and therefore be more illustrative on what kind of motion property can be matched by the proposed algorithm?

    • For example, to the best knowledge of this reviewer, there are [motion direction, speed, acceleration, spatial extent, motion periodicity].
    • The overall experimental results are very aggregated, which blocks the deeper understanding. Could benefit from category-wise result investigation.
4 comments
AuthorsOpenReview ↗

Response to Reviewer t5qF (1/2).

W0. Clarification of potential overlap between SAM 3 training data and evaluation datasets.

Answer to W0: To the best of our knowledge, the SAM 3 model was not pre-trained on the test set of MOVE/A2D dataset. First, the original datasets of MOVE and A2D are not explicitly listed among the SA-Co training sources in the SAM 3 technical report. Moreover, our experiments demonstrate that SAM 3 falls short on M-FSVOS, with a J&F score of only 30.6 on MOVE and 18.6 on A2D-Motion, as SAM 3 is not trained for motion-centric tasks. The two observations indicate that SAM 3 was not pre-trained on the test set of MOVE/A2D dataset.

W1. / Q4.2. Results on different types of motion.

Answer to W1 / Q4.2: Thank you for this insightful comment. To clarify which motion types benefit from ReMoM and how its behavior differs from SAM 3 baseline, we categorize MOVE’s 224 action classes into eight semantic subgroups within its four original domains, and report the segmentation performance of each subgroup in the table below.

DomainMotion subgroupSAM 3 RawReMoMΔ\Delta
DailyactionCommunicative hand/arm gestures40.868.1+27.3
DailyactionOther everyday activities and interactions25.057.7+32.7
EntertainmentMusical-instrument performance33.362.7+29.4
EntertainmentOther recreational activities and performances27.875.9+48.1
SportsUpper-limb actions and object manipulation25.778.8+53.1
SportsWhole-body coordination, locomotion and posture28.281.6+53.4
SpecialTransformation events34.658.7+24.1
SpecialOther special events31.344.3+13.0

As shown in the table, ReMoM substantially outperforms the SAM 3 baseline overall. It performs particularly well on whole-body coordination, locomotion and posture (81.6%) and upper-limb actions and object manipulation (78.8%). Matching these motions across support and query videos often requires stronger motion understanding because they often involve larger and more complex movements. These substantial improvements support the effectiveness of ReMoM's stronger motion understanding ability.

W2. Implications for motion-aware foundation model training.

Answer to W2: We thank the reviewer for this insightful and valuable suggestion. Our experiments demonstrate that, despite its strong object discovery and tracking capabilities, SAM 3 remains limited in understanding support-defined motions. We therefore believe that incorporating motion-centric supervision into foundation model pretraining is a promising direction, including aligning tracks that exhibit the same motion across different object appearances, distinguishing temporally coherent sequences from disrupted ones, and modeling interactions among objects. Such supervision could encourage representations that capture motion correspondence alongside object identity. We will add this discussion in the revision.

AuthorsOpenReview ↗

Response to Reviewer t5qF (2/2).

Q1. How is the number of support ways determined?

Answer to Q1: Following the standard protocol of MOVE (Ying et al., 2025), WW denotes the number of support-defined motion categories per episode. We evaluate W=2W=2 and W=5W=5 on MOVE, with one support example per category K=1K=1.

Q2. The intuition of choosing TCN.

Answer to Q2: To answer this question, we conduct an additional ablation experiment and discuss the advantages of TCN. As shown below, the Transformer-based variant achieves 63.8 J&F, compared with 68.2 for TCN. We attribute this advantage to TCN’s local temporal inductive bias, which facilitates modeling local temporal changes for motion understanding.

Motion encoderJ&F
Transformer63.8
TCN68.2

Q3. Temporal shuffling assumption.

Answer to Q3: Our temporal shuffling is useful for learning temporal structure across different types of motions, including symmetric motions. Specifically, we contrast the factual query with an appearance-preserving shuffled view, and use their relation difference to strengthen temporally sensitive motion evidence. For symmetric actions like jumping jacks, random shuffling still disrupts the coordinated transitions between successive arm and leg poses. Thus, for symmetric actions, the contrast between factual and shuffled views also encourages the model to capture temporal information.

Q4.1 Definition of motion matching.

Answer to Q4.1: In our work, motion matching refers to matching action semantics between query motions and support motions. Analogous to synonym matching, motions match when their poses, temporal progression, or effects jointly convey the same action meaning, despite differences in object identity or execution style. These cues are interpreted jointly in the context of the support-defined action category, and similarity in any single cue does not necessarily imply a match. Regarding the listed properties such as direction and speed, current benchmarks lack corresponding annotations. These limitations and the short rebuttal period prevent a systematic investigation of their individual contributions. We will clarify this definition and leave further analysis as future work.

thanks for the rebuttal. This reviwer appreicated the efforts and clarity. All concerns are addressed. Would like to stick to weak acceptance.

Dear Reviewer t5qF,

We are thrilled to note that your concerns have been addressed. We sincerely appreciate your dedicated time and effort in providing invaluable feedback and are encouraged by your recognition.

Best

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Discussion

Dear AC and Reviewers,

We are grateful for the reviewers' engagement in a fruitful discussion and for their recognition of our work. Here we provide a brief summary for the AC and all reviewers.

The reviewers have acknowledged the strengths of our work, including the effective task reformulation (snFP, fUEU), well-motivated model design (t5qF, snFP, fUEU), strong model performance (t5qF, snFP, fUEU), and clear writing (t5qF, fUEU). During the rebuttal and discussion phases, the reviewers raised some points of concern, to which we have provided detailed point-by-point responses. Following the discussion, we have adequately addressed most concerns, and Reviewer snFP has decided to increase his rating accordingly.

Specifically, our discussions with the reviewers mainly focused on three key aspects:

  • Clarification of Technical Contributions (snFP, fUEU): We provided a detailed clarification highlighting the unique contributions of our work, which have been acknowledged by Reviewers snFP and fUEU.

  • Additional Experiments and Analysis (snFP, fUEU): We conducted additional experiments and analysis to demonstrate our model's motion understanding ability and robustness. These efforts have been acknowledged by Reviewers snFP and fUEU.

  • Implementation and Experimental Details (t5qF, snFP): We supplemented the requested experimental details and provided a thorough explanation of the implementation details of our model. These efforts have been acknowledged by Reviewer t5qF and snFP.

Once again, we thank all the reviewers and the AC for their valuable efforts in reviewing our work and helping us improve it for the benefit of the wider community. We will reflect all the clarifications, discussions, and comparisons based on the reviewers' suggestions.

Best regards,

Authors of Submission 9280

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.