On the Eigenvalues of Global Covariance Pooling for Fine-Grained Visual Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Fine-Grained Visual Categorization (FGVC) is challenging due to subtle inter-class variations.
• Global Covariance Pooling (GCP) uses second-order statistics to capture these differences.
• Previous work suggested truncating small eigenvalues in GCP improves performance on large-scale datasets.

Purpose of the Study:

• Investigate why truncating small eigenvalues hinders convergence in FGVC.
• Identify the role of small eigenvalues in extracting discriminative features.
• Propose a method to enhance the importance of small eigenvalues for FGVC.

Main Methods:

• Developed two attribution methods to visualize the importance of GCP eigenvalues.
• Proposed a novel network branch to amplify small eigenvalues within the GCP layer.
• Evaluated the approach on three fine-grained benchmarks and larger datasets.

Main Results:

• Attribution visualizations revealed small eigenvalues are crucial for class-specific feature extraction.
• The proposed network branch, amplifying small eigenvalues, achieved state-of-the-art performance for GCP methods on FGVC.
• The method demonstrated competitive results against other FGVC approaches on larger datasets.

Conclusions:

• Small eigenvalues in GCP are essential for fine-grained visual categorization, contrary to previous assumptions.
• Amplifying small eigenvalues offers a parameter-free method to significantly boost FGVC performance.
• The findings provide new insights into representation learning for fine-grained visual tasks.