The recall rate of tiny-yolov3 with trt5 is lower than darknet


The tiny-yolov3 is trained by darknet,and I get the tensorrt engine by tensorrt c++ api, the precision is kfloat.
I test the tensorrt engine of tiny-yolov3 with 500 pictures,the engine can find 1864 objects.
However, I can get 2121 objects by runing darknet with the same picture.
How can I optimizate the engine?


Jetson TX2 R32.1

TensorRT Version: 5

Hi @qgq1991,

We recommend you to try with latest TensorRT version.

Thank you.