You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thank you for great work.
I tested SFace recognition.
Surprisingly, face_recognition_sface_2021dec_int8.onnx is slower about 50% than face_recognition_sface_2021dec.onnx.
Is it normal?
Initially I wanted to speed up this model by converting TF Lite.
I wonder if this would be effective or not.
The text was updated successfully, but these errors were encountered:
Surprisingly, face_recognition_sface_2021dec_int8.onnx is slower about 50% than face_recognition_sface_2021dec.onnx.
OpenCV dnn supports int8 quantized models but is not optimized good enough for that. I would suggest using non quantized version if it meets your need.
The int8 version can be slower because of how it processes data. Converting to TF Lite might speed things up, especially on devices that work well with TF Lite.
Thank you for great work.
I tested SFace recognition.
Surprisingly, face_recognition_sface_2021dec_int8.onnx is slower about 50% than face_recognition_sface_2021dec.onnx.
Is it normal?
Initially I wanted to speed up this model by converting TF Lite.
I wonder if this would be effective or not.
The text was updated successfully, but these errors were encountered: