Discrepancy between AMASS sequence Lengths and Indices in humanml_index.csv

[vadeli]

I ran into an issue with humanml.py when processing samples like CMU/80/80_63_poses.npy. It seems that the start and end indices in humanml_index.csv don't match the length of the sequence. For instance, CMU/80/80_63_poses.npy has 569 frames, but in humanml.py, there's a line:
pose = pose[int(start_frame*1.5):int(end_frame*1.5)]
And in humanml_index.csv, there's an entry like this:
./pose_data/CMU/80/80_63_poses.npy,516,716,000004.npy
This results in an empty pose. I'm wondering why the start and end indices in the CSV file aren't consistent with the length of the AMASS sequences.
Thanks

[luomingshuang]

Yes, I also meet this problem. I think it is an error.

[sohananisetty]

it might be due to the fact that the original humanml dataset (and I guess the csv) was in 20fps but this dataset is in 30fps—thats where the 1.5 factor comes in.

[luomingshuang]

The mocap_framerate in SMPL_HG is 60, while the mocap_frame_rate in SMPL_XG is 120.

[sohananisetty]

I meant I think they are using the original csv file (in the original humanml dataset) which had start and end durations according to 20fps.

[LinghaoChan]

I ran into an issue with humanml.py when processing samples like CMU/80/80_63_poses.npy. It seems that the start and end indices in humanml_index.csv don't match the length of the sequence. For instance, CMU/80/80_63_poses.npy has 569 frames, but in humanml.py, there's a line: pose = pose[int(start_frame*1.5):int(end_frame*1.5)] And in humanml_index.csv, there's an entry like this: ./pose_data/CMU/80/80_63_poses.npy,516,716,000004.npy This results in an empty pose. I'm wondering why the start and end indices in the CSV file aren't consistent with the length of the AMASS sequences. Thanks

@vadeli @sohananisetty @luomingshuang The issue comes from the inconsistency of SMPL-H and SMPL-X annotation in AMASS. The AMASS author team seems not to have replied to it yet. Plz refer to the issue and another issue.

[luomingshuang]

thanks for your reply. Another question is about the frame-level annotations in humanml. The released frame-level annotations
texts/body_texts/humanml and texts/hand_texts/humanml have some issues. For example, the number of frames in motion-x/texts/body_texts/humanml/000000.json is 97 (0-96). But the number of frames in motion_data/smplx_322/humanml/000000.npy is 146 (based on frame rate=120 and ex_fps=30). So I want to know how to match these two files. Or it still should be modified due to the inconsistent frame rate?
Hope your reply. Thanks.

[luomingshuang]

So can the generated humanml based on humanml_index.csv and amass/smpl-xg still be used normally? (even humanml_index.csv doesn't match well)