Julian Hau

Sadly not. For now I am using the results as they are, but I am still interested if someone comes up with a solution.

Do you get exactly the same detections? For me it sadly does not help as sigmoid is already applied in the onnx file.

You are right. That improved my results. Thanks! But my results are still about 1-2% worse compared to pytorch inference. May I ask what thresholds you use for the postprocessing?

When generating the .txt files for stuff and things, are the class-ids allowed to overlap between stuff and things? Do the class-ids of stuff in the .txt files need to...