Is darknet/yolov4-csp supported?

[vtyw]

Does this repo support yolov4-csp (scaled yolov4) from darknet?

I tried using a yolov4-csp model by following the same instructions as for yolov4. The detections appear fine in DeepStream, but when I check carefully, it's producing the same detections as the original model but with much lower confidence scores (e.g., 0.4 instead of 0.8).

[marcoslucianops]

Yes, it is supported, but the TensorRT reduces a bit the accuracy of the converted model.

[vtyw]

@marcoslucianops The accuracy drop is higher than I would have expected. I took the pretrained config and weights for darknet yolov4-csp and ran inference on a traffic video through darknet as well as through the equivalent deepstream pipeline that outputs detections. Both methods are using FP32 inference. The detections are roughly the same in a frame by frame comparison, except:

• Maybe a third of confidence scores are wildly mismatched (differ by more than 0.05)
• Probably due to the above, the DeepStream inference misses one or more objects (out of an average of ~14) in the majority of frames
• It's not uncommon for two matching objects to have more than 0.02 discrepancy in bounding box coordinates.

Do you know why the models perform so differently after TensorRT conversion? Even a normal FP32->FP16 conversion rarely exceeds 0.02 error in my experience. I thought it might be something to do with TensorRT not having native support for the Mish activation function.

[vtyw]

I checked whether I also experience the same inaccuracy with regular yolov4:

Error confidence scores can still exceed 0.05, but bbox coordinates are generally within 0.02 of their original value from darknet inference.

[marcoslucianops]

Do you know why the models perform so differently after TensorRT conversion?

I don't.

I checked whether I also experience the same inaccuracy with regular yolov4

It occurs with all converted models.