search

enazoe / yolo-tensorrt

TensorRT8.Support Yolov5n,s,m,l,x .darknet -> tensorrt. Yolov4 Yolov3 use raw darknet *.weights and *.cfg fils. If the wrapper is useful to you,please Star it.
MIT License
1.19k stars 315 forks source link

About the detetction precision #2

Closed guods closed 3 years ago

guods commented 4 years ago

@enazoe Did you evaluate the detection precision? Too much accuracy reduction compared to no acceleration.

enazoe commented 4 years ago

@guods Yes，I evaluate it my own simple dataset . It does not have significant performance reduction. But for the INT8 precision ,the calibration images best to cover all dataset . In other word, keep the sample diversity.

guods commented 4 years ago

@enazoe In order to avoid the inferences of calibration images, I detected my dataset on FP16，the detection results are also poor .

junhua-zhang commented 3 years ago

@guods @enazoe Correct cv4 act for bottleneck_csp to leaky relu, and precision problem should be solved.

guods commented 3 years ago

@junhua-zhang thanks, I would like to test it.