Fast Bird Part Localization for Fine-Grained Categorization

Yaser Souri, Shohreh Kasaei

Abstract: Incorporating precise part information has proved to be crucial in building accurate fine-grained categorization systems in recent studies. The state-of-the-art approach for part localization uses a convolutional neural network and needs thousands of forward passes of the network, which is very time consuming. In this paper, an efficient method is proposed for part localization, with only one forward pass of the network. The proposed method provides improved generalization capability, compared to the state-of-the-art, and the ability to detect multiple part instances simultaneously, without much computational overhead. Experiments on the Caltech-UCSD Birds dataset show that the proposed method, while being much faster, achieves comparable accuracy to the state-of-the-art.