Robust Fine-Grained Visual Categorization via Cyclical Attention.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Fine-grained visual categorization (FGVC) struggles with heavily occluded (HO) images, which obscure critical features.
• Current methods for handling occlusion have limitations, often discarding or artificially reconstructing occluded regions.

Purpose of the Study:

• To develop a novel approach for robust FGVC in open-world settings, specifically addressing the challenge of heavy occlusion.
• To leverage human self-regulated learning (SRL) principles to improve feature discrimination in occluded images.

Main Methods:

• Proposed a novel multi-instance model incorporating cyclical attention, inspired by SRL behavior.
• Developed a cooperative game mechanism within a feedback loop to recalibrate attention and re-evaluate image patch selection.
• Integrated the multi-instance model into a transformer architecture, creating the SRL-FGVC transformer.

Main Results:

• The SRL-FGVC transformer demonstrated superior performance compared to existing methods on six public datasets and a newly established dataset.
• The cyclical attention mechanism effectively mitigated attention bias by formulating a cooperative union between forward and feedback self-attention.
• The model showed consistent outperformance in heavy occlusion scenarios.

Conclusions:

• The proposed SRL-FGVC transformer offers a promising new direction for robust fine-grained visual categorization under challenging real-world conditions.
• Leveraging SRL principles and cyclical attention provides an effective strategy for handling heavily occluded image data.