STARS: Spatial Temporal Graph Convolution Network for Action Recognition System on FPGAs
Songwen Pei, Xianrong Wang, Wei Qin, Sheng Liang · 2021
Graph convolution neural network is one of the hot demanding research areas in the last few years. Due to the issues of data irregularity and computation complexity driven by typical GNN networks, we propose a spatial temporal graph convolution network for action recognition system on FPGA(STARS). STARS has redesigned several computing kernels based on the original layers of ST-GCN and adopted specific algorithm with optimization strategies for different kernels. To optimize the performance of accelerator, the ping-pong buffers for data transmission and dynamic quantification for model inference are implemented. The effectiveness of STARS driven accelerator is verified on Xilinx Pynq-Z1 prototyping board.