Clean Sample Selection and Noisy Sample Rematching for Text-Based Pedestrian Retrieval


Daiqiang Li (Sichuan University), Weicheng Zhang (Sichuan University), yuanyuan wu (Chengdu University of Technology), Honggang Chen (Sichuan University)
The 35th British Machine Vision Conference

Abstract

Text-Based Pedestrian Retrieval (TBPR) aims to match pedestrian images with corresponding text descriptions. Although notable advancements have been made in this field, most existing methods assume that training samples are clean. However, the neglect of noisy samples can degrade model performance. To address this issue, we propose an effective framework called CSNR from the perspective of dynamic dataset reconstruction, incorporating three key innovations. The Clean Sample Selection Module (CSSM) exploits the differences in the loss values of clean and noisy samples, filtering out noisy samples to improve the quality of training samples. The Noisy Sample Rematching Module (NSRM) replaces the text descriptions of some noisy samples with better-matching descriptions to create new clean samples, thereby increasing the set of usable clean data. The Weighted topK Contrastive Loss (WKCL) uses more negative samples and prioritizes harder negatives to fully leverage the information from negative samples and ensure the efficiency of model training under noisy description circumstances. Experimental results demonstrate that our approach achieves state-of-the-art performance.

Citation

@inproceedings{Li_2025_BMVC,
author    = {Daiqiang Li and Weicheng Zhang and yuanyuan wu and Honggang Chen},
title     = {Clean Sample Selection and Noisy Sample Rematching for Text-Based Pedestrian Retrieval},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_955/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection