retinaface output shape

[HYL-Dave]

output shape 是 (1, 20, 80, 80)请问这有无问题？

[zheshipinyinMc]

input shape:(1, 3, 640, 640)
strid cls bbox landmark
8 (1, 4, 80, 80) (1, 8, 80, 80) (1, 20, 80, 80)
16 (1, 4, 40, 40) (1, 8, 40, 40) (1, 20, 40, 40)
32 (1, 4, 20, 20) (1, 8, 20, 20) (1, 20, 20, 20)

[HYL-Dave]

抱歉，所以想使用ONNX的模型推论还需要对输出进行其他实作吗?还是可以在onnxruntime设置参数做推论?

[zheshipinyinMc]

onnx模型检测还需要加上后处理，参考mxnet模型的后处理

[mohamadHN93]

would you say what are the post processing to extract the landmarks or the bbox of each face in an image

[zheshipinyinMc]

post-process is convert model output to detect information，like bbox、score 、landmark

[mohamadHN93]

post-process is convert model output to detect information，like bbox、score 、landmark

Thanks for answering but could you provide a piece of code or the procedure to perform aforementioned post processing in order to extract landmarks , bbox, etc. I am working on utilizing the onnx version into tensorflow then exctracting the landmark.
I'd be grateful if you help me.
regards
or how to perform on these outputs:
input shape:(1, 3, 640, 640)
strid cls bbox landmark
8 (1, 4, 80, 80) (1, 8, 80, 80) (1, 20, 80, 80)
16 (1, 4, 40, 40) (1, 8, 40, 40) (1, 20, 40, 40)
32 (1, 4, 20, 20) (1, 8, 20, 20) (1, 20, 20, 20)