Matt Deitke Profile picture
AI Researcher at @allen_ai and @uwcse
Mar 22, 2023 14 tweets 9 min read
Introducing Objaverse, a massive open dataset of text-paired 3D objects!

Nearly 1 million annotated 3D objects to pave the way to build incredible large-scale 3D generative models: 🧵👇

🤗 Hugging Face: huggingface.co/datasets/allen…
📝ArXiv: arxiv.org/abs/2212.08051

#CVPR2023 In the past, ShapeNet has enabled remarkable progress and benchmarking across 3D computer vision!

But, it lacks visual diversity, realism…
Jun 15, 2022 9 tweets 4 min read
@andrey_kurenkov Hey Andrey! Glad to hear your excitement :)

Your comment would some nice side effects (e.g., making it more differentiable), but I disagree with using generated floorplans and object placement for this paper, and think it would actually prohibit scaling dramatically. @andrey_kurenkov For object placement generation:
- The training data isn’t that useful, since it doesn’t match the assets in the asset database (e.g., you cannot place baseball bats or basketballs if they don’t exist in any of the training scenes).