Cross-Domain Hallucination Network for Fine-Grained Object Recognition

Jin-Fu Lin, Yen-Liang Lin, Erh-Kan King, Hung-Ting Su, Winston H. Hsu; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 1214-1221

Existing fine-grained object recognition methods often require high-resolution images to better discriminate the subordinate classes. However, this assumption does not always hold in current surveillance systems, where the distinguished parts may not be clearly presented. Besides, data insufficiency and class imbalance make the problem even more challenging. In this paper, we leverage high-resolution images collected from Internet to improve the vehicle recognition in the surveillance environments. A cross-domain hallucination network is proposed to minimize the domain discrepancy and enhance the quality of low-resolution surveillance images. To better align the cross-domain features and boost the recognition performance, we extend the original framework to part-based hallucination networks, where the parts are automatically extracted based on the maximum responses from the convolution filters. Whole and part-based hallucination networks are fused in a late fusion scheme to improve the final performance. We evaluate our method on a public surveillance vehicle dataset (BoxCars21k). Experimental results demonstrate that our approach outperforms the state-of-the-art methods.