DI2L: Cross-modality person re-identification with discriminative feature and information-balanced identity learning

doi:10.1016/j.neucom.2025.132256

DI2L: Cross-modality person re-identification with discriminative feature and information-balanced identity learning

2025-12-06

0

PRE

AI

DOI：10.1016/j.neucom.2025.132256

Original

Original request for help

Save

Abstract

En 中文

• To avoid overdependence of the model on the visible modality, the latest studies usually convert RGB images to grayscale images or leverage GANs to generate auxiliary images. Different from these, we introduce a random channel augmentation (RCA) module to the visible modality instead of dropping color cues. • The latest methods usually focus only on the most discriminative rather than the diverse parts, which help distinguish different persons. Different from these, we design a non-local attention strategy with a weighted part aggregation (DPA) module to thoroughly extract identity information and abstract discriminative features. • We propose an information-balanced identity learning (I2L) module, which combines local attentive focal identity loss and a weighted regularization triplet loss (WRT) to address the information imbalance between different modalities. Compared to the commonly used hard triplet loss, WRT loss samples triplets consider their contributions, improving the robustness against modality variations. • Comprehensive experiments against the SOTA methods on three datasets (i.e., SYSU-MM01, RegDB, and LLCM) demonstrate the superiority and effectiveness of DI2L.

DI2L: Cross-modality person re-identification with discriminative feature and information-balanced identity learning

Abstract

Journal

Organization