This multiple-classes detection demo implements the lightweight Mobilenet v2 SSD network on Xilinx SoC platforms without pruning. You can learn more about mobilenetv2-SSD here.

It uses Mobilenetv2 as the backbone to significantly reduce the computational workload, which is 6.3 GOPS per image compared to 117 GOPS per image required by VGG16-SSD (both resolutions are 480*360). The network is trained based on BDD100K data with 11 classes and can achieve 20+ FPS on a low-end ZU2EG device (B1152@430MHz), achieving 40.5% efficiency.

A snapshot of the demo in action is shown below:

 

The mobilenet v2 SSD example is available in the DNNDK v3.0 package.