search

supernotman / RetinaFace_Pytorch

Reimplement RetinaFace with Pytorch
305 stars 68 forks source link

FPS of the model and small face #7

Closed luyao777 closed 5 years ago

luyao777 commented 5 years ago

Thanks for your work! Is there any speed test(FPS) of your model? thx!

luyao777 commented 5 years ago

And small face is difficult to detect, such as 0_Parade_marchingband_1_188 and 0_Parade_marchingband_1_465.

supernotman commented 5 years ago

And small face is difficult to detect, such as 0_Parade_marchingband_1_188 and 0_Parade_marchingband_1_465.

Speed has not been tested yet.Maybe in the future. And for small faces, you can change size of input from default 640 to a larger scale. Simply change detect.py, line 77: resized_img = resize(padded_img.float(),(640,640))

clancylian commented 5 years ago

Thanks for your work! Is there any speed test(FPS) of your model? thx!

hi, you will see speed test in my github, i have converted the model to caffe and then to tensorrt int8 model, you can test it! https://github.com/clancylian/retinaface

luyao777 commented 5 years ago

Thx, I tested the FPS, almost 26 FPS in widerface val for (640,640).