akanazawa/human_dynamics

resluts are wrong

Closed this issue · 3 comments

the paramers predicted are totally wrong,
图片,
what's the problem of this? can you tell me?

Hi,

This seems very weird. I'm not sure what happened to the input image either. It seems like others were able to run our demo.. so I'm not sure what is happening here. Please make sure that all the package versions match exactly whats in the README and the requirements.txt.

Best,

Angjoo

Hi,
Thanks. When the sequence length is equal to 13, the answer seems to be right. So only one frame is valid at a time.

I've face the same problem and try to reduce the sequence length equal to 13 and it doesn't work but when i try to decrease the mini-batch size from 8 to 1. It work well. What is this size of mini-batch stand for? A input image for prediction?