sirui-xu/stars

[ECCV 2022 oral] Official PyTorch implementation of the paper "Diverse Human Motion Prediction Guided by Multi-Level Spatial-Temporal Anchors"

stars's Issues

Can't replicate the results

Hello,
running the code with the default settings i can't replicate the results of the given pretrain.
Have you changed some parameters in order to obtain SotA results with the deterministic model?
The default epochs number is 50, but i suppose you have increased the number of epochs.
Thanks a lot!

Detail of 'cfg.nk' and 'stat' in 'diverse/main.py'

There have some questions about the correlation between 'nk' and ‘stat'：
’‘’
for i in range(cfg.nk):
stat[i] = (indices == i).sum()
stat /= stat.sum()
‘’‘
what's the meaning of above code in 'diverse/main.py' lines 29~31?

problem for data

Hello, Sirui-Xu.
The Link of "Human3.6m in exponential map can be downloaded from here." which in STARS/deterministic/ cannot be opend, can you give a update one?
Thanks you and Best wish.

Variation in results of various actions

Your work is adorable! I have a general question. In the results table, why does a single model result not outperform in all 15 cases of the H3.6M dataset or any other dataset?
Here, your model outperforms most cases, but in some cases, the results of the previous approach are better than this model. What is the reason behind it?
Also, in some actions, this model's results are far better than previous results, but in other actions, it is slightly better than the previous result. What is the reason for this?
Your answer will be valuable to me
Thank you!