Scaling data sets plays a pivotal role in training modern large-scale, data-driven models. While 2D vision models have benefited enormously from data scaling, progress in 3D vision models remains stunted due to the challenge of acquiring high-quality 3D data.
Addressing this disparity, a new paper titled “Objaverse-XL: A Universe of 10M+ 3D Objects” is presented by a collaborative team from the Allen Institute for AI, University of Washington, Columbia University, Stability AI, California Institute of Technology, and LAION. They introduce Objaverse-XL, a large-scale dataset of 3D assets obtained from web-crawling. With its considerable variety and quality, Objaverse-XL aims to enhance the performance of state-of-the-art 3D models significantly.
Objaverse-XL consists of a ginormous collection of diverse 3D objects in a wide range of sources, object shapes, and categories, spanning GitHub, Thingiverse, Polycam, the Smithsonian Institution, and Sketchfab. It is two orders of magnitude larger than ShapeNet – a baseline testbed for 3D objects modelling.
The data sources that has been used to create Objaverse-XL have either a strict NSFW policy or strong self-filtering, the researchers perform careful analysis and manual inspection before integrating them to Objaverse-XL, therefore ensure the high quality of Objaverse-XL.
Objaverse-XL has great potential to promise exciting novel applications in various fields, such as computer vision, graphics, augmented reality and generative AI. It paves the way to reach new levels of performance and demonstrates strong zero-shot learning capability to new domains.
In their empirical study, the team conducted zero-shot novel view synthesis with Zero123-XL and PixelNeRF. Objaverse-XL shows strong zero-shot generalization abilities and 3D models demonstrate the improvements enabled by using Objaverse-XL. The team hopes that Objaverse-XL can serve as a foundation for future work in 3D.
The paper Objaverse-XL: A Universe of 10M+ 3D Objects on arXiv.
Author: Hecate He | Editor: Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.