A Knowledge-Driven Enhanced Module for Visible-Infrared Person Re-Identification

EasyChair Preprint 8469

Abstract

Compared to traditional person re-identification, which only handles the intra-modality discrepancy, the Visible-Infrared Person Re-identification (VI-ReID) suffers from additional cross-modality discrepancy caused by the cross-domain inherent heterogeneity. However, most existing VI-ReID methods ignore the corresponding relationships of intrinsic property knowledge inside cross-modality. Inspired by the human brain's cognitive process of knowledge, in this paper, a novel Knowledge-driven Enhance Module (KDEM) method is designed to imitate the cognitive process of the human brain to achieve the effective matching of cross modalities. Our proposed KDEM aims to discover and integrate the significant semantic pattern from cross-modality representations into a new knowledge-enhanced modality and further enhance the matching accuracy of cross modalities. Meanwhile, a diversity loss is designed to exclude redundant knowledge and preserve the variety of semantic knowledge in the integrated knowledge-enhanced modality. Moreover, a consistency loss is designed to preserve the semantic correlation between the integrated knowledge-enhanced modality and the other two modalities. The evaluation results on two popular benchmark datasets demonstrated the effectiveness of the proposed KDEM, and it obtained competitive performance compared to state-of-the-art methods on the VI-ReID task. The source code of our KDEM is released at https://github.com/SWU-CS-MediaLab/KDEM.

Keyphrases: Deep Neural Network, cross-modality retrieval, person re-identification

@booklet{EasyChair:8469,
  author    = {Shihao Shan and Enyuan Xiong and Xiang Yuan and Song Wu},
  title     = {A Knowledge-Driven Enhanced Module for Visible-Infrared Person Re-Identification},
  howpublished = {EasyChair Preprint 8469},
  year      = {EasyChair, 2022}}