search

Gaoyiminggithub / Graphonomy

Graphonomy: Universal Human Parsing via Graph Transfer Learning
MIT License
295 stars 68 forks source link

Why the speed is much slow than SSL #5

Closed wugh closed 5 years ago

wugh commented 5 years ago

Hi, Gaoyiminggithub

I'm look for a good human semantic parser recently I found serval projects.

Because the LIP_SSL didn't annotated the neck part of human, but your code can handle this. But why the speed of Graphonomy is much more slower than LIP_SSL, is there anyway I can speed up.

Gaoyiminggithub commented 5 years ago

Hi, There may be two reasons for that: 1. SSL tests only use one scale while we use 7 scales. You can change the scale list to [1] in here to speed up. 2. SSL uses the framework of Caffe which may be faster than Pytorch.

wugh commented 5 years ago

Thanks, I will try the scale first, I also see some related section in your paper. But there is no related experiments to evaluate the performance if we inference without multi scale.

Gaoyiminggithub commented 5 years ago

Hi, the reason why there are no experiments without multi-scale is that previous work reported the results with multi-scale.