The Whole Is More Than Its Parts? From Explicit to Implicit Pose Normalization

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Fine-grained classification involves recognizing visually similar object categories, such as bird species.
• Traditional methods relied on explicit pose normalization, which involves detecting and describing object parts.
• Recent approaches using global average or bilinear pooling achieve comparable accuracy without explicit pose normalization.

Purpose of the Study:

• To analyze the advantages of global average and bilinear pooling over generic Convolutional Neural Networks (CNNs) and explicit pose normalization.
• To demonstrate how these pooling methods achieve implicit object pose normalization.
• To introduce novel visualization techniques for understanding pose handling in CNNs and improving model interpretability.

Main Methods:

• Introduced 'activation flow,' a novel visualization technique to investigate pose handling limitations in traditional CNNs (e.g., AlexNet, VGG).
• Presented and compared 'neural activation constellations' for explicit pose normalization.
• Proposed 'α-pooling,' a generalized framework for global average and bilinear pooling.

Main Results:

• α-pooling achieved higher accuracy, improving common CNN models by up to 22.9 percent.
• Explicit pose normalization approaches offered better interpretability compared to pooling methods.
• A visualization approach was developed to enhance the understanding and analysis of model predictions.

Conclusions:

• Global average and bilinear pooling methods offer significant accuracy improvements in fine-grained classification.
• Novel visualization techniques aid in understanding and addressing limitations in pose handling within CNNs.
• The proposed approaches demonstrate benefits for fine-grained recognition and other fields like action recognition.