search

cvlab-columbia / zero123

Zero-1-to-3: Zero-shot One Image to 3D Object (ICCV 2023)
https://zero123.cs.columbia.edu/
MIT License
2.61k stars 188 forks source link

Where is valid_paths.json #22

Closed wzic closed 1 year ago

wzic commented 1 year ago

Hi! Happy to read about your excellent work!

May I know where is the file valid_paths.json that is used for training on objaverse? I can only find object-paths.json in the downloaded files.

Thanks.

ruoshiliu commented 1 year ago

Hi @wzic , a few hundreds (or thousands) of our downloaded files are corrupted or didn't render correctly, so I created a valid_paths.json to store the path to valid rendering folders.

wzic commented 1 year ago

Then what's the content in valid_paths.json? Is it a list of the ids of the objects? How about the format?

ojmichel commented 1 year ago

@wzic From what I can tell it is just a list of the valid object uids in json

buttomnutstoast commented 1 year ago

@ruoshiliu , it'd be very helpful if you can release the valid_paths.json file, otherwise, we are not able to know which objects are used for training/validation.

ruoshiliu commented 1 year ago

valid_paths.json.zip There you go! The validation is the last 1 percent of the rows (first 99 percent used for training).

P.S. if you want to compare against our model on objaverse dataset, please use 105000.ckpt instead of 165000.ckpt as the latter might have been trained on some part of the validation sets unintentionally. As our paper focuses on zero-shot generalization, it didn't really matter but it does when you run in-distribution experiments.

VitorGuizilini-TRI commented 1 year ago

Hi, can you explain in more details how the evaluation is performed? More specifically, how do you choose which image you use as context, and which ones you use for novel view synthesis?

ruoshiliu commented 1 year ago

For GSO and RTMV, we render a bunch of views whose camera poses are randomly sampled. We use the first view as input and the following views for evaluation. Since all views are uniformly sampled, the order doesn't make a difference. Same applies for Objaverse but since our paper focuses on zero-shot performance, we did not run evaluation on objaverse which is our training dataset.