Objaverse: A Universe of Annotated 3D Objects

Recent dramatic advances in artificial intelligence have been driven by massive web-scale datasets for text and 2D images. In contrast, 3D computer vision and embodied AI have remained constrained by small, mid-sized repositories that lack visual diversity and category breadth. This article introduces Objaverse 1.0, an open-access repository containing over 800,000 richly annotated 3D objects created by more than 150,000 artists, to demonstrate how scaling 3D data enables new capabilities across generative modeling, 2D recognition, robotic simulation, and model evaluation.

The authors curated the dataset from the online platform Sketchfab under Creative Commons licenses, organizing foundational metadata such as descriptions, tags, animations, and articulable parts. To support specific vision benchmarks, they constructed Objaverse-LVIS, a curated subset of 47,000 objects mapped across 1,156 granular semantic categories using automated text-image filtering and crowdsourced human verification. The researchers evaluated the utility of this resource across four distinct application domains: training 3D generative meshes, augmenting long-tailed 2D instance segmentation, building open-vocabulary robotic navigation environments, and benchmarking model robustness against non-standard viewing angles.

The evaluations yielded four primary findings. First, in 3D generative modeling, human evaluators judged objects produced by models trained on Objaverse to be more visually diverse than those trained on previous datasets in 91% of comparisons. Second, using rendered 3D objects as synthetic data augmentations improved 2D instance segmentation performance across rare, common, and frequent object categories. Third, integrating Objaverse into simulated home environments scaled the variety of unique objects from around 2,000 to 36,000 across 1,100 categories, enabling autonomous agents to successfully navigate to open-vocabulary target descriptions with a 19.9% success rate compared to 5.1% for random baselines. Fourth, evaluating leading vision-language models on objects viewed from random orientations exposed major vulnerabilities, as accuracy dropped by roughly 18 to 23 percentage points when moving from canonical views to arbitrary camera angles.

These findings indicate that large-scale 3D data provides a cost-effective path to improving both 2D and 3D computer vision systems. By substituting expensive manual annotations with synthetic renderings, organizations can enhance recognition performance on rare real-world categories and safely test safety-critical models against perspective shifts before physical deployment. Furthermore, the findings show that existing standard vision models overfit heavily to canonical forward-facing viewpoints, highlighting an unaddressed operational risk for autonomous systems operating in unpredictable physical environments.

Organizations developing computer vision and autonomous systems should integrate diverse 3D asset pipelines to augment low-sample training data and establish mandatory multi-angle robustness evaluations. Teams should also utilize scaled 3D environments to expand the vocabulary and generalizability of embodied AI agents. Future work should expand fine-grained category annotations across the remaining uncurated assets, address unnatural scaling or noise in web-scraped metadata, and develop targeted training interventions to close the substantial performance gaps identified in non-canonical viewpoint recognition.