1000fps human segmentation with deep convolutional neural networks

Efficiency and effectiveness are two key factors to evaluate a human segmentation algorithm for real vision applications. However, most existing algorithms only focus on one of them. That is, fast and accurate human segmentation is not yet well addressed. In this paper, we propose a super-fast and highly accurate human segmentation method with very deep convolutional neural networks. We also provide a comprehensive study on the proposed approach, including different net structures, various techniques of alleviating over-fitting, and performance enhancement with different extra data. Experimental results on the database of Baidu people segmentation competition [1] demonstrate that the proposed model outperforms traditional segmentation algorithms in accuracy and speed. Although it is slightly worse than the very complex champion algorithm, it is encouraging that our method can obtain more than 10,000 times acceleration, showing that it has great potential for practical applications.