Learning Compact Identity Representations for Weakly Textured Hanwoo Cattle Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animal Science

Background:

• Cross-view re-identification of Hanwoo cattle is difficult due to weak textures and significant pose/viewpoint variations.
• These variations cause feature dispersion, scattering samples of the same individual and creating ambiguous local neighborhoods.

Purpose of the Study:

• To develop a framework that explicitly improves identity compactness for weakly textured cattle ReID.
• To address challenges posed by pose and viewpoint changes in cattle re-identification.

Main Methods:

• Proposed a framework utilizing Pose- and Text-Conditioned Inpainting Diffusion (PTID) for generating pose-diverse, identity-consistent samples.
• Introduced Dual-Adaptive Viewpoint-Aware Feature Centralization (DVFC) to refine features by exploiting viewpoint consistency and neighborhood reliability.
• Constructed a new Hanwoo ReID dataset with 37 identities across diverse farm environments.

Main Results:

• The framework significantly reduced pose-induced feature dispersion.
• Achieved substantial improvements in re-identification accuracy, with gains up to 40.8 mAP and 26.1 Rank-1.
• Demonstrated consistent performance improvements on multiple baselines in both closed-set and open-set settings.

Conclusions:

• Explicitly improving identity compactness is a viable strategy for robust cattle ReID.
• The proposed PTID and DVFC methods offer a promising solution for weakly textured cattle re-identification challenges.
• The developed dataset facilitates further research in this domain.