search

facebookresearch / VidOSC

Code and data release for the paper "Learning Object State Changes in Videos: An Open-World Perspective" (CVPR 2024)
Other
29 stars 1 forks source link

Missing .npz files in /videoclip_pseudolabel #10

Open profundus59 opened 1 week ago

profundus59 commented 1 week ago

Hi,

Thank you for sharing your great work! I am facing a minor issue while running the evaluation code of HowToChange with the given checkpoints: python train.py --det 0 --sc_list all --ckpt ./checkpoints/multitask.ckpt

When I run this, I'm getting this warning in my terminal: SCR-20241002-lbhj

The reproduced results are okay (+- 1~2% from the paper), but I wanted to check if you are aware of these missing files.

Thanks!

profundus59 commented 1 week ago

The videoclip_pseudolable files I used are from the link in the repository: https://drive.google.com/file/d/1YQn7KeRduXeMABBLAP43uNqz8RpOTrI1/view

profundus59 commented 6 days ago

Also taking a look at the .npz files from videoclip_pseudolabel, about 50 ~ 80% of the annotation for the frames are done as "-1", which refers "ambiguous" if my understanding is correct.

Looking at it, would it be fair to think that VideoCLIP is incapable of doing the proper job of pseudo-labeling frames based on the three state descriptions for each OSC?

I'm curious what is your take on this. Thanks!