Egocentric Voxel Lifting (EVL)

Overview

EVL is an opensource AI model from Meta for 3D object detection and surface reconstruction on ProjectAria recordings. EVL is trained on simulation data, namely Aria Synthetic Environments. It relies on frozen 2D foundation features to set a competitive performance. EVL leverages all egocentric modalities from Aria including posed and calibrated RGB and greyscale video streams and semidense points.

Further reading

• Research Paper - EFM3D: A Benchmark for Measuring Progress Towards 3D Egocentric Foundation Models
• Github repo github.com/facebookresearch/efm3d

Getting Started

To start using the model, check out the EVL model inference demo in the EFM3D repo.