ROMP icon indicating copy to clipboard operation
ROMP copied to clipboard

h36m processed frame number mismatch

Open miracleyoo opened this issue 2 years ago • 2 comments

https://github.com/Arthur151/ROMP/blob/db7e2a1df1df104909e6b137eaaed653ee433517/romp/lib/dataset/preprocess/h36m_extract_frames.py#L40

Hi there, Thanks for the great work! I'm trying to train the model on the h36m dataset, and I downloaded the original h36m dataset to process using this script. However, the generated frame number is mismatched with the one in your processed h36m dataset.

Firstly, this dex<20 condition seems not really being applied when you processed the data, otherwise, each video will have at most 20 frames.

Then, even if I removed this condition, the generated frame number for each video is still different from your processed h36m data (on google drive). For example, the video Purchases.60457274.mp4 for S1 generates 231 frames (index from 0-230) using this script (after moving the dex<20 condition). However, in the google drive processed version, there are only 230 frames (index 0-229). For reference, in the original 50 FPS video, the lasting time is 23.2s, and the frame number is 1161.

The similar thing happened for other videos as well, but the delta number is not always 1. For instance, the S1's WalkTogether.60457274.mp4 video has a 2 frame difference (245(your processed) vs 247(actual)).

Since I'm trying to train an event-camera time surface-based model that has a high requirement for time accuracy, I hope you can help clarify where is this missing frame, should I drop some frame at the beginning, or drop the last K frame, or somewhere in the middle. If possible, could you please share the exact processing script you used for generating the current processed h36m dataset? I really appreciate it.

miracleyoo avatar Mar 11 '23 00:03 miracleyoo

@miracleyoo Really sorry for the inconvienence this problem caused.

I can't find the exact code for extracting the frame. This code is provided by other developers who use ROMP.

I just remember that I dropped some frames at the beginning of each sequence.

Maybe you can extract all frames of a video and match them with the processed h36m frames to get the precise frame time.

Arthur151 avatar Mar 16 '23 11:03 Arthur151

@miracleyoo

in case zou have calculated the smpl models for h36m. could you tell me, which camera parameters did you utilize for further usage. Are they the original camera parmaters or some other as per ROMP.

I am waiting for your response

Dipankar1997161 avatar Jun 09 '23 16:06 Dipankar1997161