Yolo (You Only Look Once) is a very popular CNN model for object detection. It has several versions, with the latest Yolov3 having the best accuracy. We have included an ADAS detection demo using Yolov3 trained with the Cityscapes dataset in the Xilinx DNNDK v2.08 download available here.
Below is a screenshot from the demo.
Yolov3 is based on the Darknet Framework. To run it on Xilinx devices, we provide a reference design which includes a conversion tool that converts the original Darknet model to a Caffe model. The tool can also test the detection result of both Darknet and Caffe model to provide an accuracy comparison. Once the model accuracy is acceptable, Xilinx's DNNDK tool chain is used to perform quantization, compilation and deployment to the ZCU102 board.
The directory structure of the package is shown as follows,
To find out more, please contact your Xilinx representative.