Deep CNNs Meet Global Covariance Pooling: Better Representation and Generalization

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Existing deep convolutional neural networks (CNNs) often use global average pooling, which may not fully capture deep feature statistics.
• Global covariance pooling offers potential for improved representation and generalization but faces challenges in robust estimation and geometric utilization of high-dimensional, small-sample deep features.

Purpose of the Study:

• To propose a novel global covariance pooling method, Matrix Power Normalized COVariance (MPN-COV), to address limitations of existing pooling techniques in deep CNNs.
• To enhance the representation and generalization abilities of deep CNNs by effectively utilizing rich feature statistics.

Main Methods:

• Developed Matrix Power Normalized COVariance (MPN-COV) pooling, a robust covariance estimator suitable for high-dimensional, small-sample scenarios, leveraging the geometry of covariances via a Power-Euclidean metric.
• Introduced a global Gaussian embedding network to integrate first-order statistics with MPN-COV.
• Implemented iterative matrix square root normalization for efficient training, avoiding eigen-decomposition, and utilized progressive 1x1 convolutions and group convolutions for covariance representation compression.

Main Results:

• MPN-COV demonstrates superior performance in large-scale object classification, scene categorization, fine-grained visual recognition, and texture classification compared to existing methods.
• The proposed methods achieve state-of-the-art results across various visual recognition tasks.
• The modular design allows easy integration into existing deep CNN architectures.

Conclusions:

• The MPN-COV pooling method effectively addresses the challenges of robust covariance estimation and geometric utilization in deep CNNs.
• The proposed approach significantly improves the representation and generalization capabilities of deep CNNs, leading to state-of-the-art performance on diverse visual recognition benchmarks.
• MPN-COV offers a practical and effective enhancement for deep learning models in computer vision.