SthPhoenix / InsightFace-REST

InsightFace REST API for easy deployment of face recognition services with TensorRT in Docker.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Accuracy loss ?

MyraBaba opened this issue · comments

Hi,

Is there any accuracy decrease after conversion from standard model ?

I have noticed that ArcFace outputs are a bit different from MXNet model, though very close, similarity between MXNet embeddings and converted ONNX embeddings is somewhere around 0.999998.

Outputs of ONNX and TRT models are exactly the same.

I haven't run any benchmarks on accuracy though, I'll appreciate if someone could check this.

Closing in favor of #34