Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamic Uncertainty Learning with Noisy Correspondence for Text-Based Person Search

Zequn Xie, Haoming Ji, Chengxuan Li, Lingwei Meng

TL;DR

The Dynamic Uncertainty with Noisy Correspondences DUNC framework is proposed, which introduces a novel Cross-modal Uncertainty Learning paradigm and a robust loss function, Dynamic Robust Loss DRL, which enhances robustness and representation quality, even in noisy environments.

Abstract

Text-to-image person search aims to identify an individual based on a text description. To reduce data collection costs, large-scale text-image datasets are created from co-occurrence pairs found online. However, this can introduce noise, particularly mismatched pairs, which degrade retrieval performance. Existing methods often focus on negative samples, which amplify this noise. To address these issues, we propose the Dynamic Uncertainty and Relational Alignment (DURA) framework, which includes the Key Feature Selector (KFS) and a new loss function, Dynamic Softmax Hinge Loss (DSH-Loss). KFS captures and models noise uncertainty, improving retrieval reliability. The bidirectional evidence from cross-modal similarity is modeled as a Dirichlet distribution, enhancing adaptability to noisy data. DSH adjusts the difficulty of negative samples to improve robustness in noisy environments. Our experiments on three datasets show that the method offers strong noise resistance and improves retrieval performance in both low- and high-noise scenarios.

Dynamic Uncertainty Learning with Noisy Correspondence for Text-Based Person Search

TL;DR

The Dynamic Uncertainty with Noisy Correspondences DUNC framework is proposed, which introduces a novel Cross-modal Uncertainty Learning paradigm and a robust loss function, Dynamic Robust Loss DRL, which enhances robustness and representation quality, even in noisy environments.

Abstract

Text-to-image person search aims to identify an individual based on a text description. To reduce data collection costs, large-scale text-image datasets are created from co-occurrence pairs found online. However, this can introduce noise, particularly mismatched pairs, which degrade retrieval performance. Existing methods often focus on negative samples, which amplify this noise. To address these issues, we propose the Dynamic Uncertainty and Relational Alignment (DURA) framework, which includes the Key Feature Selector (KFS) and a new loss function, Dynamic Softmax Hinge Loss (DSH-Loss). KFS captures and models noise uncertainty, improving retrieval reliability. The bidirectional evidence from cross-modal similarity is modeled as a Dirichlet distribution, enhancing adaptability to noisy data. DSH adjusts the difficulty of negative samples to improve robustness in noisy environments. Our experiments on three datasets show that the method offers strong noise resistance and improves retrieval performance in both low- and high-noise scenarios.
Paper Structure (22 sections, 15 equations, 3 figures, 2 tables)

This paper contains 22 sections, 15 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Text-to-Image Person Search
  4. Uncertainty-based Learning
  5. Methodology
  6. Feature Extraction with Dual-Encodert
  7. Key Feature Selector
  8. Uncertainty Modeling
  9. Cross-modal Evidential Learning
  10. Dynamic Softmax Hinge Loss
  11. Cross-Modal Alignment
  12. Overall Loss Function
  13. Experiments
  14. Datasets and Performance Measurements
  15. Implementation Details
  16. ...and 7 more sections

Figures (3)

  • Figure 1: The overview of our Dynamic Uncertainty and Relational Alignment framework (DURA).illustrating feature extraction, evidence-based uncertainty modeling, and evidence-guided training with TAL, CEL, DSH, and KFS.
  • Figure 2: We visualize the evidence distribution of clean and noisy pairs at different training stages of our DURA, which is conducted on CHUKPEDES under 20% noise.
  • Figure 3: Test performance (Rank-1) versus epochs on the CHUKPEDES and ICFG-PEDES datasets with 50% noise.