Facebook AI Research releases Common Objects in 3D Dataset(CO3D)

Facebook AI Research, FAIR has released the Common Objects in 3D, CO3D dataset. A dataset of 1.5 million frames from 19,000 videos of real-world objects along with their highly accurate 3D reconstructions spread across 50 distinct categories. This makes it the first of its kind, and largest dataset, both by categories and objects compared to its alternatives. Facebook AI Research informed about this development in one of their latest tweets.

Reconstructing objects in 3D has many applications in the field of Computer Vision. These applications range from using reconstructions of objects to blend virtual and real objects for the purpose of immersive AR/ VR applications on phones, laptops, or other devices generation of 3D models in gaming. Current methods use learning models to reconstruct 3D models. This approach is however hindered by the lack of datasets that contain both, videos of real-world and accurate reconstructions of them. To train the model, these reconstructions are created with the help of synthetic objects. These synthetic objects are only able to approximate the actual objects and are hence not accurate. To help bridge this gap between the need for more accurate reconstructions of objects and the lack of this data, Facebook AI Research has released the Common Objects in 3D, CO3D Dataset.

Besides, the dataset can also be used for benchmarking the methods used for the reconstruction of 3D objects such as Implicit Differentiable Renderer, Neural Radiance Fields- NeRF, etc.

The Facebook AI Research Team has also used the dataset to train NeRFormer- a deep learning network that learns the geometric structure of object categories by observing the videos provided for training.

You can check out the CO3D Dataset over here.


Check out the blog post on Computer Vision Trends to look out for in 2021, here.

Check out the blog post on 4 Trends in Machine Learning to watch out for, here.