Understanding deep features with computer-generated imagery

TL;DR

This paper presents a method to analyze CNN features by controlling scene factors through computer-generated imagery, revealing how different scene attributes influence network responses across various CNN architectures.

Contribution

The study introduces a novel approach using rendered images to dissect CNN feature responses with respect to scene factors, providing insights into network behavior and differences across architectures.

Findings

CNN responses vary significantly with scene factors

Different CNN architectures respond uniquely to scene attributes

Analysis of computer-generated images translates to natural image representations

Abstract

We introduce an approach for analyzing the variation of features generated by convolutional neural networks (CNNs) with respect to scene factors that occur in natural images. Such factors may include object style, 3D viewpoint, color, and scene lighting configuration. Our approach analyzes CNN feature responses corresponding to different scene factors by controlling for them via rendering using a large database of 3D CAD models. The rendered images are presented to a trained CNN and responses for different layers are studied with respect to the input scene factors. We perform a decomposition of the responses based on knowledge of the input scene factors and analyze the resulting components. In particular, we quantify their relative importance in the CNN responses and visualize them using principal component analysis. We show qualitative and quantitative results of our study on three CNNs trained on large image datasets: AlexNet, Places, and Oxford VGG. We observe important differences across the networks and CNN layers for different scene factors and object categories. Finally, we demonstrate that our analysis based on computer-generated imagery translates to the network representation of natural images.