Open Deante-dx opened 2 months ago
If you want to use incremental information as model input, you should add the following two lines of code before line 52 in /utils/h36motion3d.py
:
pad_idx = np.repeat([input_n - 2], output_n+1)
i_idx = np.append(np.arange(0, input_n-1), pad_idx)
Here, the new i_idx
will mask the incremental information of the last frame of historical motion, which is the correct approach.
: (
Thank you for your reminding. We will check and deal with it immediately. I hope the research content in other parts of the paper can be helpful to your research.
We have checked this issue and found that the implementation in our code indeed implicitly included information about the first frame of future motion. We converted all motion sequences to increment sequences to speed up the training process, but forgot to exclud the increment vector of the last frame in each historical sequence, which was derived based on the last frame of the historical sequence and the first frame in the future sequence, resulting in the mistake. We are re-implementing all related experiments. We will correct the results, report this case to ACM MM23, and ask for withdrawn in the next few days. I will update the progress to you. I sincerely apologize for any inconvenience and misleading this may have caused to you, other readers, and our community. Thank you very much for informing us this issue!
Thank you 👍 ^_^
When using incremental information, you include information about the first frame of future human motion. You can directly obtain the true value of the first frame of future human motion by using the last frame of historical incremental information combined with the last frame of historical human motion. Therefore, there is an error in your training process. After fixing this error, the prediction performance will be poor.