search

omniobject3d / OmniObject3D

[ CVPR 2023 Award Candidate ] OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation
462 stars 12 forks source link

Questions about dataset statistics #5

Closed TLB-MISS closed 1 year ago

TLB-MISS commented 1 year ago

First of all, thank you so much for providing such a great dataset. I have minor questions about dataset statistics. The paper states that the dataset provides a total of 6000 instances in 190 categories. However, when I downloaded the dataset, there were a total of 5385 instances in 216 categories. Having problems with my download? Or is it correct that the number of instances is less than the paper?

Thank you

wutong16 commented 1 year ago

Hi, thank you for your attention to the paper.

Regarding the instance number, as stated in the readme, we have reserved 10% of the instances as a "hidden" test set for an upcoming challenge focused on sparse-view nerf and 3D generation. We plan to release the codebase and available sparse views for these test sets later this month. However, the final version of the test data may not be as large as initially indicated. The remaining instances will be merged with the current training set and released too in this month.

In terms of the category number, it is indeed subject to change due to recent efforts to divide certain categories into more fine-grained subcategories.

TLB-MISS commented 1 year ago

@wutong16 Ah, sorry. I should have read the documentation more carefully. Thanks for the kind reply!