A Semantic-Aware Attention and Visual Shielding Network for Cloth-Changing Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Cloth-changing person re-identification (ReID) is challenging due to appearance variations caused by different clothing.
• Existing ReID methods often overlook human semantic information and feature consistency across clothing changes.
• Focusing solely on body shape or contour sketches limits the robustness of feature representations.

Purpose of the Study:

• To propose a novel semantic-aware attention and visual shielding network (SAVS) for robust cloth-changing person ReID.
• To address the limitations of current methods by focusing on clothing-invariant semantic features.
• To develop a unified framework that jointly utilizes human semantic attention and visual clothes shielding.

Main Methods:

• A visual semantic encoder is utilized for human body and clothing region localization via semantic segmentation.
• A human semantic attention (HSA) module is designed to emphasize semantic information and reweight feature maps.
• A visual clothes shielding (VCS) module is developed to cover clothing regions, focusing on invariant semantic features.

Main Results:

• The proposed SAVS method significantly outperforms state-of-the-art approaches on cloth-changing person ReID tasks.
• SAVS achieves substantial improvements in mean average precision (mAP) and rank-1 accuracy on benchmark datasets like LTCC and Celeb-reID.
• Compared to MBUNet and Swin Transformer, SAVS demonstrates superior performance across multiple datasets.

Conclusions:

• The SAVS network effectively extracts robust features for cloth-changing person ReID by focusing on semantic information.
• The proposed approach successfully overcomes the challenges posed by appearance variations due to clothing changes.
• This work provides a significant advancement in person re-identification research, particularly for scenarios involving altered attire.