A Lightweight Network Based on Pyramid Residual Module for Human Pose Estimation

The human pose estimation is one of the most popular research fields. Its current accuracy is satisfactory in some cases, however, there exists a challenge for practical application due to the limited memory and computational efficiency in FPGAs and other hardware. We propose a lightweight module based on the pyramid residual module in this work. We change the convolution mode by using the depth-wise separable convolutions structure. Meanwhile, the channel split module and channel shuffle module are added to change the feature graph dimension. As a result, the parameters of the network are reduced effectively. We test the network on standard benchmarks MPII dataset, our method reduces about 50% of the training storage space while maintaining comparable accuracy. The complexity is simplified from 9 GFLOPs to 3 GFLOPs.