Interactive attack-defense for generalized person re-identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generalized Person Re-Identification (GReID) seeks models that generalize to new domains with limited training data.
• Attack-Defense mechanisms are used for GReID, treating domain shifts as attacks and employing defense modules.
• Existing methods struggle with complex domain transformations due to separated attack and defense components.

Purpose of the Study:

• To introduce an innovative Interactive Attack-Defense (IAD) mechanism for GReID.
• To overcome limitations of separated attack-defense components in existing GReID approaches.
• To enhance the generalization capabilities of GReID models across unseen domains.

Main Methods:

• Developed an Interactive Attack-Defense (IAD) mechanism with mutually learning attack and defense models.
• The attack model generates dynamic, directional attacks based on the defense model's state.
• The defense model learns generalizable representations from diverse attack samples through a dual training process.

Main Results:

• The IAD mechanism demonstrated superior performance compared to state-of-the-art methods in GReID.
• Achieved a 13.4% improvement in average Rank-1 accuracy on the MSMT17 dataset under various protocols.
• Experimental results across multiple datasets confirm the effectiveness of the proposed IAD approach.

Conclusions:

• The interactive learning framework of IAD fosters synergistic learning between attack and defense models.
• IAD effectively models complex domain transformations, leading to robust GReID performance.
• The proposed method significantly advances the state-of-the-art in Generalized Person Re-Identification.