Fine-Grained Categorization by Deep Part-Collaboration Convolution Net

Abstract

In part-based categorization context, the ability to learn representative feature from quantitative tiny object parts is of similar importance as to exactly localize the parts. We propose a new deep net structure for fine-grained categorization that follows the taxonomy workflow, which makes it interpretable and understandable for humans. By training customized sub-nets on each manually annotated parts, we increased the state-of-the-art part-based classification accuracy for general fine-grained CUB-200-2011 dataset by 2.1%. Our study shows the proposed method can produce more activation to discriminate detail part difference while maintaining high computing performance by applying a set of strategies to optimize the deep net structure.