Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SAS-VPReID: A Scale-Adaptive Framework with Shape Priors for Video-based Person Re-Identification at Extreme Far Distances

Qiwei Yang, Pingping Zhang, Yuhao Wang, Zijing Gong

TL;DR

SAS-VPReID targets video-based ReID under extreme far-distance conditions, introducing a triad of modules: MEVB to stabilize CLIP-based features with video-consistent augmentation and multi-proxy memory, MGTM to capture multi-granularity temporal cues with learnable scale fusion, and PRSD to incorporate clothing-invariant 3D shape dynamics via SMPL priors and Transformer-based temporal modeling. The framework is trained with a combination of discriminative, contrastive, and shape-regularization losses, and evaluated on DetReIDXV1, where it achieves state-of-the-art results across aerial-ground and cross-session settings, including notable gains in A→G, G→A, and A→A scenarios. Key findings include the effectiveness of memory-based supervision for extreme degradation, the value of adaptive temporal fusion for tracklets of varying quality, and the benefit of explicit SMPL shape priors to reduce clothing-induced appearance drift. The work demonstrates practical impact by delivering robust identity representations under challenging conditions and offering a scalable approach that can leverage larger backbones for further gains without sacrificing stability.

Abstract

Video-based Person Re-IDentification (VPReID) aims to retrieve the same person from videos captured by non-overlapping cameras. At extreme far distances, VPReID is highly challenging due to severe resolution degradation, drastic viewpoint variation and inevitable appearance noise. To address these issues, we propose a Scale-Adaptive framework with Shape Priors for VPReID, named SAS-VPReID. The framework is built upon three complementary modules. First, we deploy a Memory-Enhanced Visual Backbone (MEVB) to extract discriminative feature representations, which leverages the CLIP vision encoder and multi-proxy memory. Second, we propose a Multi-Granularity Temporal Modeling (MGTM) to construct sequences at multiple temporal granularities and adaptively emphasize motion cues across scales. Third, we incorporate Prior-Regularized Shape Dynamics (PRSD) to capture body structure dynamics. With these modules, our framework can obtain more discriminative feature representations. Experiments on the VReID-XFD benchmark demonstrate the effectiveness of each module and our final framework ranks the first on the VReID-XFD challenge leaderboard. The source code is available at https://github.com/YangQiWei3/SAS-VPReID.

SAS-VPReID: A Scale-Adaptive Framework with Shape Priors for Video-based Person Re-Identification at Extreme Far Distances

TL;DR

SAS-VPReID targets video-based ReID under extreme far-distance conditions, introducing a triad of modules: MEVB to stabilize CLIP-based features with video-consistent augmentation and multi-proxy memory, MGTM to capture multi-granularity temporal cues with learnable scale fusion, and PRSD to incorporate clothing-invariant 3D shape dynamics via SMPL priors and Transformer-based temporal modeling. The framework is trained with a combination of discriminative, contrastive, and shape-regularization losses, and evaluated on DetReIDXV1, where it achieves state-of-the-art results across aerial-ground and cross-session settings, including notable gains in A→G, G→A, and A→A scenarios. Key findings include the effectiveness of memory-based supervision for extreme degradation, the value of adaptive temporal fusion for tracklets of varying quality, and the benefit of explicit SMPL shape priors to reduce clothing-induced appearance drift. The work demonstrates practical impact by delivering robust identity representations under challenging conditions and offering a scalable approach that can leverage larger backbones for further gains without sacrificing stability.

Abstract

Video-based Person Re-IDentification (VPReID) aims to retrieve the same person from videos captured by non-overlapping cameras. At extreme far distances, VPReID is highly challenging due to severe resolution degradation, drastic viewpoint variation and inevitable appearance noise. To address these issues, we propose a Scale-Adaptive framework with Shape Priors for VPReID, named SAS-VPReID. The framework is built upon three complementary modules. First, we deploy a Memory-Enhanced Visual Backbone (MEVB) to extract discriminative feature representations, which leverages the CLIP vision encoder and multi-proxy memory. Second, we propose a Multi-Granularity Temporal Modeling (MGTM) to construct sequences at multiple temporal granularities and adaptively emphasize motion cues across scales. Third, we incorporate Prior-Regularized Shape Dynamics (PRSD) to capture body structure dynamics. With these modules, our framework can obtain more discriminative feature representations. Experiments on the VReID-XFD benchmark demonstrate the effectiveness of each module and our final framework ranks the first on the VReID-XFD challenge leaderboard. The source code is available at https://github.com/YangQiWei3/SAS-VPReID.
Paper Structure (21 sections, 20 equations, 2 figures, 5 tables)

This paper contains 21 sections, 20 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Video-based Person Re-Identification
  4. Cloth-changing Person Re-Identification
  5. Aerial-Ground Person Re-Identification
  6. Methodology
  7. Task Definition
  8. Memory-Enhanced Visual Backbone
  9. Multi-Granularity Temporal Modeling
  10. Prior-Regularized Shape Dynamics
  11. Training and Inference
  12. Experimental Setups
  13. Datasets and Evaluation Protocols
  14. Implementation Details
  15. Experimental Results
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Examples illustrating four major variations in the VReID-XFD challenge: (a) distance variation (different distances leading to noticeable size and clarity differences); (b) viewpoint variation (frontal, side, top-down and tilted views); (c) outfit variation (outfit changes across sessions causing large appearance drifts) and (d) occlusion variation (occlusions and similar-texture backgrounds increasing confusion).
  • Figure 2: Illustration of our proposed SAS-VPReID framework for the VReID-XFD Challenge. The Memory-Enhanced Visual Backbone (MEVB) strengthens the visual feature representation, with video-consistent data augmentation and memory-based supervision to stabilize the model adaptation under extreme far-distance degradation. Then, the Multi-Granularity Temporal Modeling (MGTM) captures multi-granularity spatiotemporal features via efficient sequence modeling and learnable scale fusion. Finally, the Prior-Regularized Shape Dynamic (PRSD) injects clothing-invariant identity cues by modeling SMPL-based shape parameters over time, with an explicit shape prior and temporal aggregation. With the integration of MEVB, MGTM and PRSD, our SAS-VPReID effectively learns discriminative multi-granularity spatiotemporal and clothing-invariant features for extreme far-distance video-based person ReID.