search

mks0601 / PoseFix_RELEASE

Official TensorFlow implementation of "PoseFix: Model-agnostic General Human Pose Refinement Network", CVPR 2019
MIT License
327 stars 64 forks source link

speed of PoseFix #16

Closed thancaocuong closed 5 years ago

thancaocuong commented 5 years ago

thank you for that awesome release. But can you tell me how to calculate speed per one image when testing?

mks0601 commented 5 years ago

can you clarify what you want to know? do you want to know how to measure seconds in python?

thancaocuong commented 5 years ago

I want to calculate FPS when testing on COCO val2017. I ran your test script on COCO val2017 and got same AR and AP as SOTA. But the total time was 7h27m54s. I ran on single GTX 1080 TI. I don't know why it took so long.

mks0601 commented 5 years ago

54th line of test.py is start of each cropped image's testing. You should start to measure second at line 55 and ends at line 128. The indentation level of line 55 and 128 should be the same. I also feel that running time of PoseFix is slow. I think that is because input is not RGB, but RGB + heatmap. As heatmap is very large tensor (384x288xjoint_num), it will take time.

thancaocuong commented 5 years ago

thank you so much. I will try this ^^