search

vye16 / slahmr

MIT License
441 stars 50 forks source link

Error when running long sequence #17

Closed Coral79 closed 1 year ago

Coral79 commented 1 year ago

Hi, I had some problems, the code crushed when I tried to run 1000 frames of my video. I saw the default number of input frames is 120. I am wondering if you have tried to run the code on a longer sequence.

Thank you very much!

geopavlakos commented 1 year ago

The current optimization schedule behaves better for shorter sequences (e.g., up to 200-300 frames) and we have observed difficulties in optimizing longer sequences. It would probably be preferable to split the sequences in smaller batches first and run slahmr on each one of them.

Coral79 commented 1 year ago

Hi, thank you very much for the reply. I have noticed for the egobody dataset, you split them into chunks (100 framse). In this case, since the video is optimized in small batches, will each chunk have a different global scale and ground postion?

geopavlakos commented 1 year ago

Yes, if the subsequences are optimized independently, the global scale and the ground plane will be different.