Unified Multi-Dataset Training for TBPS

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chatterjee, Nilanjana, Garg, Sidharatha, Subramanyam, A V, Lall, Brejesh
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917215398264832
author Chatterjee, Nilanjana
Garg, Sidharatha
Subramanyam, A V
Lall, Brejesh
author_facet Chatterjee, Nilanjana
Garg, Sidharatha
Subramanyam, A V
Lall, Brejesh
contents Text-Based Person Search (TBPS) has seen significant progress with vision-language models (VLMs), yet it remains constrained by limited training data and the fact that VLMs are not inherently pre-trained for pedestrian-centric recognition. Existing TBPS methods therefore rely on dataset-centric fine-tuning to handle distribution shift, resulting in multiple independently trained models for different datasets. While synthetic data can increase the scale needed to fine-tune VLMs, it does not eliminate dataset-specific adaptation. This motivates a fundamental question: can we train a single unified TBPS model across multiple datasets? We show that naive joint training over all datasets remains sub-optimal because current training paradigms do not scale to a large number of unique person identities and are vulnerable to noisy image-text pairs. To address these challenges, we propose Scale-TBPS with two contributions: (i) a noise-aware unified dataset curation strategy that cohesively merges diverse TBPS datasets; and (ii) a scalable discriminative identity learning framework that remains effective under a large number of unique identities. Extensive experiments on CUHK-PEDES, ICFG-PEDES, RSTPReid, IIITD-20K, and UFine6926 demonstrate that a single Scale-TBPS model outperforms dataset-centric optimized models and naive joint training.
format Preprint
id arxiv_https___arxiv_org_abs_2601_14978
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Unified Multi-Dataset Training for TBPS
Chatterjee, Nilanjana
Garg, Sidharatha
Subramanyam, A V
Lall, Brejesh
Computer Vision and Pattern Recognition
Text-Based Person Search (TBPS) has seen significant progress with vision-language models (VLMs), yet it remains constrained by limited training data and the fact that VLMs are not inherently pre-trained for pedestrian-centric recognition. Existing TBPS methods therefore rely on dataset-centric fine-tuning to handle distribution shift, resulting in multiple independently trained models for different datasets. While synthetic data can increase the scale needed to fine-tune VLMs, it does not eliminate dataset-specific adaptation. This motivates a fundamental question: can we train a single unified TBPS model across multiple datasets? We show that naive joint training over all datasets remains sub-optimal because current training paradigms do not scale to a large number of unique person identities and are vulnerable to noisy image-text pairs. To address these challenges, we propose Scale-TBPS with two contributions: (i) a noise-aware unified dataset curation strategy that cohesively merges diverse TBPS datasets; and (ii) a scalable discriminative identity learning framework that remains effective under a large number of unique identities. Extensive experiments on CUHK-PEDES, ICFG-PEDES, RSTPReid, IIITD-20K, and UFine6926 demonstrate that a single Scale-TBPS model outperforms dataset-centric optimized models and naive joint training.
title Unified Multi-Dataset Training for TBPS
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2601.14978