Unsupervised Learning of Intrinsic Semantics with Diffusion Model for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised person re-identification (Re-ID) aims to retrieve persons without identity labels.
• Existing methods often struggle with noise and lack of semantic consistency between patch and global features.

Purpose of the Study:

• To propose a Person Intrinsic Semantic Learning (PISL) framework using a diffusion model for unsupervised person Re-ID.
• To improve the discriminative semantic structure and semantic consistency in Re-ID.

Main Methods:

• Developed a Spatial Diffusion Model (SDM) for sampling patches with intrinsic semantic structure.
• Introduced Semantic Controlled Diffusion (SCD) loss to generate semantic patches.
• Proposed Patch Semantic Consistency (PSC) loss to refine global features and pseudo-labels.

Main Results:

• The PISL framework demonstrated superior performance compared to existing unsupervised Re-ID methods.
• Experiments on three challenging datasets validated the effectiveness of the proposed approach.

Conclusions:

• The proposed diffusion model-based framework effectively addresses limitations in unsupervised person Re-ID.
• PISL framework achieves state-of-the-art results by enhancing semantic patch generation and feature consistency.