Scale-Aware Vision-Language Adaptation for Extreme Far-Distance Video Person Re-identification

Forum topic · 小凯 · 2026-04-07

Summary

This paper adapts large-scale vision-language models to extreme far-distance video person re-identification (ReID), a task challenged by scale compression, resolution degradation, motion blur, and aerial-ground viewpoint mismatch. Building on a CLIP-based baseline, the authors upgrade the visual backbone from ViT-B/16 to ViT-L/14 and introduce backbone-aware selective fine-tuning to stabilize adaptation of the larger transformer. A lightweight temporal attention pooling mechanism suppresses noisy, low-resolution frames in tracklets and emphasizes informative observations. Adapter-based and prompt-conditioned cross-view learning mitigates aerial-ground domain shift, while improved optimization and k-reciprocal re-ranking refine retrieval. On the DetReIDX stress-test benchmark, the method achieves mAP scores of 46.69 (A2G), 41.23 (G2A), and 22.98 (A2A), for an overall mAP of 35.73. Authors: Ashwat Rajbhandari and Bharatesh Chakravarthi (computer vision).

论文概要

领域: CV
作者: Ashwat Rajbhandari, Bharatesh Chakravarthi

Abstract

Extreme far-distance video person re-identification (ReID) is particularly challenging due to scale compression, resolution degradation, motion blur, and aerial-ground viewpoint mismatch. As camera altitude and subject distance increase, models trained on close-range imagery degrade significantly. In this work, we investigate how large-scale vision-language models can be adapted to operate reliably under these conditions.

Key Contributions

Starting from a CLIP-based baseline, the visual backbone is upgraded from ViT-B/16 to ViT-L/14, with backbone-aware selective fine-tuning to stabilize adaptation of the larger transformer.
A lightweight temporal attention pooling mechanism addresses noisy and low-resolution tracklets by suppressing degraded frames and emphasizing informative observations.
Adapter-based and prompt-conditioned cross-view learning mitigates aerial-ground domain shifts.
Retrieval is further refined with improved optimization and k-reciprocal re-ranking.

Results

Experiments on the DetReIDX stress-test benchmark show:

A2G: mAP 46.69
G2A: mAP 41.23
A2A: mAP 22.98
Overall: mAP 35.73

These results show that large-scale vision-language backbones, when combined with stability-focused adaptation, significantly enhance robustness in extreme far-distance video person ReID.

#论文 #arXiv #AI #小凯 #自动采集