Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

S3-CLIP: Video Super Resolution for Person-ReID

Tamas Endrei, Gyorgy Cserey

TL;DR

S3-CLIP introduces a GAN-free, task-driven video super-resolution pipeline designed to boost cross-view person re-identification under severe resolution gaps. By coupling a SwinIR-based SR module with CLIP-based ReID in a two-phase training regime and enforcing temporal consistency, the approach improves tracklet quality without adversarial training. Evaluated on DetReIDX, it yields substantial gains in ground-to-aerial matching, while remaining competitive in aerial-to-ground and aerial-to-aerial settings, highlighting the value of SR as a preprocessing step for ReID in challenging deployments. The work also emphasizes limitations such as extreme degradations and aspect-ratio distortions, pointing to future directions in multi-scale SR and authentic degradation modeling to enhance real-world robustness.

Abstract

Tracklet quality is often treated as an afterthought in most person re-identification (ReID) methods, with the majority of research presenting architectural modifications to foundational models. Such approaches neglect an important limitation, posing challenges when deploying ReID systems in real-world, difficult scenarios. In this paper, we introduce S3-CLIP, a video super-resolution-based CLIP-ReID framework developed for the VReID-XFD challenge at WACV 2026. The proposed method integrates recent advances in super-resolution networks with task-driven super-resolution pipelines, adapting them to the video-based person re-identification setting. To the best of our knowledge, this work represents the first systematic investigation of video super-resolution as a means of enhancing tracklet quality for person ReID, particularly under challenging cross-view conditions. Experimental results demonstrate performance competitive with the baseline, achieving 37.52% mAP in aerial-to-ground and 29.16% mAP in ground-to-aerial scenarios. In the ground-to-aerial setting, S3-CLIP achieves substantial gains in ranking accuracy, improving Rank-1, Rank-5, and Rank-10 performance by 11.24%, 13.48%, and 17.98%, respectively.

S3-CLIP: Video Super Resolution for Person-ReID

TL;DR

S3-CLIP introduces a GAN-free, task-driven video super-resolution pipeline designed to boost cross-view person re-identification under severe resolution gaps. By coupling a SwinIR-based SR module with CLIP-based ReID in a two-phase training regime and enforcing temporal consistency, the approach improves tracklet quality without adversarial training. Evaluated on DetReIDX, it yields substantial gains in ground-to-aerial matching, while remaining competitive in aerial-to-ground and aerial-to-aerial settings, highlighting the value of SR as a preprocessing step for ReID in challenging deployments. The work also emphasizes limitations such as extreme degradations and aspect-ratio distortions, pointing to future directions in multi-scale SR and authentic degradation modeling to enhance real-world robustness.

Abstract

Tracklet quality is often treated as an afterthought in most person re-identification (ReID) methods, with the majority of research presenting architectural modifications to foundational models. Such approaches neglect an important limitation, posing challenges when deploying ReID systems in real-world, difficult scenarios. In this paper, we introduce S3-CLIP, a video super-resolution-based CLIP-ReID framework developed for the VReID-XFD challenge at WACV 2026. The proposed method integrates recent advances in super-resolution networks with task-driven super-resolution pipelines, adapting them to the video-based person re-identification setting. To the best of our knowledge, this work represents the first systematic investigation of video super-resolution as a means of enhancing tracklet quality for person ReID, particularly under challenging cross-view conditions. Experimental results demonstrate performance competitive with the baseline, achieving 37.52% mAP in aerial-to-ground and 29.16% mAP in ground-to-aerial scenarios. In the ground-to-aerial setting, S3-CLIP achieves substantial gains in ranking accuracy, improving Rank-1, Rank-5, and Rank-10 performance by 11.24%, 13.48%, and 17.98%, respectively.
Paper Structure (17 sections, 12 equations, 2 figures, 1 table)

This paper contains 17 sections, 12 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related works
  3. Proposed Method
  4. Baseline algorithm
  5. CLIP-ReID
  6. Video Set-Level-Adapter CLIP
  7. Sampling for Super Resolution ReID Training
  8. Video Super Resolution Loss
  9. Experiments
  10. Dataset.
  11. Evaluation protocol.
  12. Network Architecture.
  13. Implementation details.
  14. Quantitative Results
  15. Fail cases
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Overview of the S3-CLIP architecture during the second-stage of training. Two types of tracklets are sampled according to SING: (i) a high-resolution tracklet that is synthetically downscaled, and (ii) a naturally low-resolution tracklet of the same person. These low-resolution tracklets are concatenated and fed into the super-resolution network. The outputs are then bicubically upscaled to match the input size of the CLIP-based visual encoder. A two-phase training strategy is employed: first, only the super-resolution network is updated while the visual encoder remains frozen; then, the super-resolution network is frozen and the visual encoder is updated.
  • Figure 2: Representative failure cases where super-resolution degrades ReID performance due to extremely low resolution, aspect ratio mismatch, JPG compression and motion blur.