Find Hidden Modality Divergence: Adversarial Aware Learning for Unsupervised Visible-Infrared Person...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised visible-infrared person re-identification (VI-ReID) faces challenges due to significant modality gaps.
• Existing contrastive learning methods struggle with modality divergence, hindering cross-modality feature learning.

Purpose of the Study:

• To propose a novel framework, Adversarial Aware Learning (ADAL), to mitigate modality divergence in unsupervised VI-ReID.
• To enhance the discriminative power of identity features across visible and infrared modalities.

Main Methods:

• ADAL explores and adversarially optimizes instances that cause modal divergence.
• It facilitates positive optimization directions for cluster centroids while penalizing negative ones.
• Instance-level optimization increases affinities for positive pairs with large cross-modality gaps.

Main Results:

• Extensive experiments demonstrate the effectiveness of the proposed ADAL framework.
• ADAL significantly outperforms existing state-of-the-art unsupervised VI-ReID methods.
• The method acts as a universally applicable plug-in module for existing approaches.

Conclusions:

• ADAL effectively addresses the modality divergence problem in unsupervised VI-ReID.
• The proposed framework improves feature learning and person re-identification accuracy across modalities.
• ADAL offers a valuable enhancement for current unsupervised VI-ReID techniques.