The evaluation of the models employs three comprehensive real-world datasets: OmniObject3D, Ocrtoc3D, and Pix3D, ensuring a robust assessment for zero-shot generalization.
OmniObject3D presents a diverse collection of 3D scans and videos across 216 categories, facilitating training on household products, toys, and food, with improved rendering techniques.
Ocrtoc3D focuses on object-centric videos, providing detailed 3D annotations across 15 coarse categories, which enhances the dataset's relevance by ensuring quality through manual curation.
Pix3D, with its 3D annotations from 9 categories, complements the evaluation by offering object masks and CAD models, making it versatile in testing model performance.
Collection
[
|
...
]