zst123 / xilinx_kria_firai Goto Github PK

Hello!

We study your project in tutorial and try to implement the same yolov4 model by our own dataset. After iterative training, the model can display a good quality in server.

However, when we use the Vitis-AI to convert and quantize the model's weights, and deploy it on the KV260, the test script shows that the detection precision, recall and f1-score are very low. Another problem we found is the detected object seems have a lot of overlapped boxes and the boxes' size are very large. As shown in following figure.

By the way, on the board's inference, we have tried this script and your application.

We would like to know if the problem occurred in the test script (YOLOv3 and YOLOv4 test scripts do not match) or in the quantification of the model?

Any suggestions to optimize the quantization of the model?

Thanks in advance!