search

cvlab-columbia / zero123

Zero-1-to-3: Zero-shot One Image to 3D Object (ICCV 2023)
https://zero123.cs.columbia.edu/
MIT License
2.67k stars 192 forks source link

Question about `total_view` used for dataset #91

Open Xiashangning opened 1 year ago

Xiashangning commented 1 year ago

Hi, first thanks for your great project! I am trying to reproduce the results with diffusers so I have examined the code quite thoroughly. But it seems that the hyper-parameter total_view in the config file provided is inconsistent with the paper. If I understand the logic of getitem in ObjaverseData correctly, it will select randomly two images of the same object in the range [0, total_view[ but the value in the config file is just 4 instead of 12... Do you only use the first 4 camera poses for training or is it a mistake when uploading the config file?

Luh1124 commented 1 year ago

Hello, I am also trying to reproduce the results of the paper. I was wondering if you have downloaded the complete data file views_release.tar.gz and how much storage space it takes up approximately after decompression?Thanks

Xiashangning commented 1 year ago

Around 1.7T

Jason-aplp commented 9 months ago

@Xiashangning Hello, I was wondering if you have figured the problem out. But I think it may be alright to just change the value in the config file from 4 to 12 since there are 12 images for each object. And I believe this may have something to do with https://github.com/cvlab-columbia/zero123/issues/60#issue-1800595706