yolov5 and deepstream accuracy problem

[qybing]

The tensorrt weights I exported using the yolov5 official website are almost the same as the original weight prediction results, and the error range is also within an acceptable range, but the weights I derived according to your method are sometimes very different in the results predicted by Deepstream. Do you have any suggestions?

[marcoslucianops]

I'm trying to increase the accuracy of YOLOv5 on DeepStream. The current implementation is the most accurate I've gotten so far. The TensorRT layers (from DeepStream) are the same PyTorch layers, but there's an accuracy drop on DeepStream.

[Arpit1496]

I am also getting really Low Accuracy on yolov5.
Is this relates to waning I get while converting model "Warning [TRT: Subnormal FP16 values detected." ?

[marcoslucianops]

I am also getting really Low Accuracy on yolov5. Is this relates to waning I get while converting model "Warning [TRT: Subnormal FP16 values detected." ?

It's not related to that.

[Arpit1496]

okay. Any updates on this ?
How can we help to improve the accuracy ?

[marcoslucianops]

No updates yet.

[Alberto1404]

@marcoslucianops Does this happens as well in recent versions of YOLO?

[marcoslucianops]

@marcoslucianops Does this happens as well in recent versions of YOLO?

Yes, but, in my use cases, it's not a relevant drop, so I still using DeepStream and YOLO on projects.

[OctaM]

I've noticed this for yoloV8 also. Are there any updates?

[sruap1214]

I've noticed this for yolov7-tiny. I have a model for 6-classes, and one particular class has a major drop in comparisson with pytorch

[pullmyleg]

Hi @marcoslucianops we're having this issue too but are keen to help resolve it. Do you have any suggestions on where we should start?

We have also attempted to resolve this by using yolov8 as an onnx model, which had similar results.

We're seeing a significant drop in accuracy for some objects (40%) when running inference with ultralytics yolov8 vs in deep stream. With yolov4 we do not see much of a drop in accuracy comparison to running inference with darknet vs deep stream.

Any tips or comments on where we should begin would be greatly appreciated! Thanks!

[KindOfBlue7]

Is there a certain YOLO model that works fine after deployment without any issues?

[marcoslucianops]

We have also attempted to resolve this by using yolov8 as an onnx model, which had similar results.

#339

Is there a certain YOLO model that works fine after deployment without any issues?

I'm adding new benchmarks to the README.md. You will be able to see the results and choose the best option.