| Details: |
|
Date：
|
24/04/2025 (Thursday)
|
|
Time：
|
15:00-17:30
|
|
Venue：
|
LT2, Wui Chi Building, Main Campus
|
|
Student：
|
Yongheng Qian
|
|
Topic：
|
Study of Cross-Modality Person Re-Identification using Data Generation and Representation Learning
|
|
Abstract： |
With the large-scale deployment of intelligent cameras that automatically switch between visible and infrared modes under varying illumination conditions, visible-infrared crossmodality person re-identification (VI-ReID) becomes an important research branch within the computer vision community. VI-ReID aims to retrieve a person of interest across disjoint cross-modality camera views. It plays a crucial role in smart cities, public security, and national governance. However, modality differences and intra-modality variations (e.g., viewpoint, misalignment, occlusion, pose, and illumination) make the VI-ReID extremely challenging. This research project aims to apply generative modeling and representation learning paradigms to explore methods for enhancing data diversity and obtaining better abstraction representation space to address the above challenges. This thesis follows the general workflow of deep learning and provides a detailed introduction to the solutions for data preprocessing and network architecture design. These include pose attention-guided paired-images generation (PAPG) network is proposed for data augmentation. To obtain a robust discriminative representation space, a dual-space aggregation learning (DSAL) network is proposed. Subsequently, a multi-level contrastive learning network with hierarchical knowledge synergy (MCLNet) is proposed based on deep supervised learning to capture low-level modality-invariant features. To fully mine the rich visual semantics of patch tokens and deep fuse specific local features and global information, a hybrid ResNetTransformer hierarchical aggregation architecture (HAResformer) is proposed. HAResformer has the potential to become a new baseline for the VI-ReID task. |
|
Enquiry：
|
fca@mpu.edu.mo
|
|
|---|
