Win-Win by Competition: Auxiliary-Free Cloth-Changing Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (ReID) faces challenges with clothing changes (CC-ReID).
• Current CC-ReID methods often rely on auxiliary information, increasing complexity and computational cost.
• This reliance on auxiliary data limits effectiveness and scalability.

Purpose of the Study:

• To develop an effective CC-ReID method that does not require auxiliary information.
• To improve identity preservation by leveraging inherent image features.
• To enhance overall system efficiency and reduce complexity in CC-ReID.

Main Methods:

• Introduced the Auxiliary-free Competitive IDentification (ACID) model.
• Implemented a hierarchical competitive strategy for feature extraction at global, channel, and pixel levels.
• Integrated appearance and structure features, reconstructed images, and employed generative adversarial learning with self- and cross-ID penalties.

Main Results:

• The ACID model demonstrated superior performance on four public cloth-changing datasets (PRCC-ReID, VC-Cloth, LTCC-ReID, Celeb-ReID).
• Achieved significant improvements in identity preservation and reduced intra-class variations.
• Showcased the effectiveness of an auxiliary-free approach in CC-ReID.

Conclusions:

• The proposed ACID model offers a highly effective and efficient solution for cloth-changing person re-identification.
• Leveraging internal image information through competitive strategies surpasses traditional auxiliary-dependent methods.
• ACID represents a significant advancement in robust person re-identification systems.