This high-accuracy pose detection demo implements the well-known OpenPose network using Xilinx DNNDK on the ZCU102 platform. You can learn more about OpenPose here
The OpenPose model is very resource intensive. The baseline network (368 x 368-pixel input) requires 268 Gops per image! With DNNDK pruning it drops to 50Gops per image (with 81% of the network pruned away!). This allows us to achieve 20+ FPS with live USB camera input on a ZCU102 board, with the DPU operating at 77% efficiency.
The demo platform setup and detection result are shown as below.