acceptodds
Withdrawn from ICLR 2027

GaitWM: Generative World Models for Sparse-View Gait Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Gait recognition has achieved remarkable progress with the advancement of deep learning, showing promising performance in both controlled environments and unconstrained real-world scenarios. Existing cross-view gait recognition methods typically rely on densely collected multi-view gait sequences to learn view-consistent representations, assuming sufficient observations across predefined viewpoints. However, such assumptions rarely hold in practical applications, where gait data are often collected from sparse and unpredictable viewpoints due to the high cost and difficulty of multi-view acquisition and annotation. This limitation significantly degrades recognition performance when encountering unseen viewpoints. To address this challenge, we propose , utilizing a for label-efficient . Unlike conventional methods that learn view-invariant representations solely from existing observations, GaitWM learns the underlying gait dynamics and view transformation knowledge from generative video world, enabling the generation of plausible gait sequences from unobserved viewpoints. By synthesizing complementary views conditioned on sparse observations, GaitWM effectively expands viewpoint coverage and provides richer motion cues for recognition without requiring densely collected multi-view gait datasets. Extensive experiments on CASIA-B and SUSTech1K demonstrate that GaitWM consistently outperforms existing methods under sparse-view settings, achieving robust generalization to unseen viewpoints. The results validate the effectiveness of generative video models as a powerful paradigm for reducing reliance on exhaustive multi-view gait data collection and advancing practical gait recognition systems.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gait Recognition, Generative World Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FBN5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies sparse-view gait recognition, where each training identity is observed from only one viewpoint rather than the dense multi-view coverage normally assumed in cross-view gait recognition. GaitWM uses the pretrained Vista4D video-reshooting model to reconstruct a dynamic 4D scene from an observed gait sequence and generate additional gait videos from six novel azimuths. The generated sequences inherit the source identity and are used as additional training samples. To reduce the effect of identity corruption in synthetic videos, the authors introduce Generative Identity Regularization, consisting of stronger label smoothing for generated samples and an angle-dependent triplet margin. Experiments on SUSTech1K and CASIA-B show substantial improvements under the proposed sparse-view protocol. For example, BigGait improves from 45.2% to 87.1% on SUSTech1K and from 49.6% to 76.0% on CASIA-B, with gains also observed for GaitBase, DeepGaitV2, and Gait-X.

Critical strengths
  1. The empirical effect under the proposed sparse-view setting is large and consistent across multiple recognition architectures and input modalities. GaitWM improves not only BigGait but also GaitBase, DeepGaitV2, and Gait-X, making it unlikely that the result is specific to one recognizer. The appendix also reports cross-domain improvements, which strengthen the claim that the generated viewpoints provide useful supervision beyond a single benchmark.

  2. The choice of a temporally consistent 4D/video-reshooting model is well matched to gait recognition, where preserving motion dynamics is more important than independently generating plausible frames. The paper also provides useful qualitative failure analysis, including background distortion and failures at extreme viewpoint changes, rather than presenting only successful generations.

Critical weaknesses
  1. The manuscript presents GaitWM as a new paradigm for sparse-view gait recognition, but the positioning relative to the closest prior work is incomplete. GaitCSV already explicitly studies sparse-view gait recognition, and prior gait view-synthesis methods have used generated viewpoints to improve recognition. More recent generative gait-recognition methods further weaken the broad claim that generative modeling for biometric recognition is largely unexplored. Most importantly, GaitCSV is discussed in the related work but is not included in Table 1. As a result, the current experiments establish that GaitWM substantially improves conventional gait recognizers under the proposed sparse-view setting, but they do not establish superiority over the closest method specifically designed for sparse-view recognition.

  2. Section 4.1 states that the SUSTech1K training viewpoint is randomly sampled for each identity, whereas Appendix A.2 states that the assignment is deterministic and requires no random seed. This inconsistency directly affects reproducibility because the identity-to-view assignment determines the difficulty of the sparse-view training protocol. In addition, the “dense-view data + 6 synthetic views” result in Table 5 is almost numerically identical to the sparse-view BigGait+GaitWM result in Table 1, despite the former supposedly retaining the full dense-view real training set. This requires clarification because it affects the interpretation of the claimed source-domain degradation and cross-domain trade-off.

  3. The dominant improvement comes from adding synthetic viewpoints, while Generative Identity Regularization contributes only a relatively small additional gain. However, the paper does not compare Vista4D-based synthesis against a simpler geometry-based augmentation, another gait-specific view-synthesis method, or an alternative video-generation approach. Therefore, the current evidence supports the conclusion that additional synthetic viewpoints are highly beneficial, but it does not yet establish that the specific generative world-model formulation is necessary. A controlled synthesis baseline is important for attributing the gain to the proposed methodology rather than to viewpoint augmentation more generally.

Questions for authors
  1. In Table 5, why does the “dense-view data + 6 synthetic views” experiment produce a SUSTech1K result almost identical to the sparse-view BigGait+GaitWM result in Table 1? Please confirm exactly which real training sequences were included and whether this row comes from an independent training run.

  2. Can the authors provide a comparison with GaitCSV under exactly the same one-view-per-identity protocol? Since GaitCSV is explicitly designed for sparse-view gait recognition, this comparison would materially affect my assessment of the contribution.

  3. Is the SUSTech1K sparse-view assignment random, as stated in Section 4.1, or deterministic, as stated in Appendix A.2? Please provide the exact identity-to-view assignment procedure and, if randomness is involved, the seed and sensitivity across different assignments.

  4. Can the authors compare Vista4D augmentation against a simpler view-generation baseline or an existing gait-specific view-synthesis approach? This would clarify whether the large gain is specifically attributable to the proposed generative video/world-model approach rather than to synthetic viewpoint augmentation in general.

Reviewer_nCutOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GaitWM for sparse-view gait recognition by leveraging a pretrained video generation model to synthesize gait sequences from unobserved viewpoints. The generated samples are used for data augmentation, while Generative Identity Regularization adjusts identity supervision to account for potential generation errors. Experiments on CASIA-B and SUSTech1K demonstrate recognition improvements under sparse-view training settings.

Critical strengths
  1. The proposed approach achieves substantial accuracy gains under sparse-view settings, indicating the effectiveness of generated multi-view samples for gait recognition.

  2. The approach is evaluated on multiple gait recognition models, including GaitBase, DeepGaitV2, BigGait, and Gait-X, showing consistent improvements across different input modalities.

Critical weaknesses
  1. The framework primarily adopts the pretrained Vista4D model to generate additional training samples, without developing a new world model or integrating generative modeling into the recognition framework. Moreover, Generative Identity Regularization mainly adapts existing label-smoothing and triplet losses, offering limited methodological innovation.

  2. Although GaitCSV (ACM MM 2023) is discussed in Related Work, no direct experimental comparison is provided. Evaluating GaitCSV under the same sparse-view protocol, or explaining potential protocol incompatibilities, would strengthen the evaluation.

  3. The experiments mainly demonstrate the benefits of adding generated views, without comparing simpler viewpoint augmentation or transformation methods. Such comparisons are necessary to establish whether large-scale video generation is essential.

  4. Although the paper evaluates identity consistency through feature similarity and recognition accuracy, these evaluations do not directly verify whether fine-grained gait dynamics are preserved during generation. Quantitative analysis of motion consistency would strengthen the conclusions.

  5. The proposed approach requires expensive video generation, with approximately 3,840 GPU hours and $32K reported in Section 4.5. The practical cost-effectiveness of this approach remains unclear, particularly compared with lightweight data augmentation methods.

Questions for authors

See Weaknesses

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

This paper presents GaitWM, a framework for sparse-view gait recognition that uses a pretrained video world model to generate gait sequences from novel viewpoints. The generated sequences are used to supplement limited cross-view training data. The authors also introduce Generative Identity Regularization to account for potential identity inconsistencies in generated samples. Experiments on SUSTech1K and CASIA-B show substantial improvements across several existing gait recognition models.

Critical strengths
  1. The paper addresses a practical problem. Collecting dense multi-view gait data is expensive, and existing recognition methods perform poorly when only limited viewpoints are available for training.
  2. The experimental results are strong. GaitWM substantially improves recognition accuracy on both datasets, particularly with BigGait and Gait-X. The improvements across different recognition backbones are also encouraging.
Critical weaknesses
  1. The motivation for using a world model is not fully convincing. The authors argue that 4D-aware generation can better preserve gait dynamics than skeleton- or SMPL-based methods. However, it is not clear why a world model is particularly necessary for this task, rather than other view synthesis approaches. As a result, the method feels somewhat driven by the availability of a powerful pretrained model rather than a clearly established task-specific need.
  2. The experimental comparisons are insufficient. The paper shows that adding generated views improves recognition performance, but does not compare against alternative view generation methods or existing methods specifically designed for sparse-view gait recognition, such as GaitCSV and GaitSSB. This makes it difficult to assess the specific contribution of the world-model-based design.
Questions for authors
  1. Could the authors clarify what makes a world model particularly suitable for sparse-view gait recognition? In particular, what advantages does the 4D scene representation provide for preserving identity-related gait dynamics compared with other generation approaches?
  2. Could the authors provide comparisons with alternative view synthesis methods, such as skeleton- or SMPL-based generation, under the same training protocol? It would also be helpful to compare with existing sparse-view gait recognition methods. These results would help determine whether the improvements come mainly from adding synthetic views or from the specific choice of a world model.

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

Related papers

Discussion (0)

Sign in to comment.