Efficient pose machine based on parameter-sensitive hashing

In this paper, we propose an efficient pose machine using Parameter-Sensitive Hashing(PSH) techniques. Based on the original pose machine, which is a sequential prediction framework, we employ the Convolutional Neural Network(CNN) to extract features. To handle the high dimensional feature vectors and conduct similarity search efficiently, we use the Parameter-Sensitive Hashing Function(PSHF) to map the feature vectors into binary values. The property of the PSHF ensures that the collisions happen when two vectors are near to each other and the search can be completed in a fractional power time. We apply our approach to the popular datasets including LSP and FLIC and make a comparison with previous methods based on a criterion of strict Percentage of Correct Parts(PCP). Experimental results reflect that our approach outperforms previous methods in accuracy.