Related Experiment Video
Updated: Jun 27, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.8K
Occlusion-Aware Transformer With Second-Order Attention for Person Re-Identification
Summary
This study introduces OAT, a novel method for person re-identification (ReID) that tackles occlusion challenges. OAT utilizes second-order attention and entropy-guided fusion to improve feature learning and achieve more accurate pedestrian recognition.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Person re-identification (ReID) faces significant challenges due to occlusion in real-world scenarios.
- Existing methods often compromise semantic information or increase network complexity when addressing occlusion.
Purpose of the Study:
- To propose a novel method, OAT, for robust person re-identification under occlusion.
- To enhance feature learning by capturing higher-order relations and dynamically adjusting feature importance.
Main Methods:
- Employs a Transformer backbone with multiple class tokens for diverse pedestrian feature learning.
- Introduces a Second-Order Attention (SOA) module to capture comprehensive features, with efficient approximation formulations.
- Proposes an Entropy Guided Fusion (EGF) module to dynamically weight class tokens based on information entropy, mitigating occlusion uncertainty.
Main Results:
- The proposed OAT method demonstrates effectiveness on both occluded and holistic person re-identification datasets.
- The SOA and EGF modules contribute to learning more discriminative class token representations.
Conclusions:
- OAT offers a novel and effective approach to person re-identification, particularly in the presence of occlusion.
- The integration of second-order attention and entropy-guided fusion significantly improves the robustness and accuracy of ReID systems.
