Abstract: Even when applied to 2D images, natural language describes a fundamentally 3D world. We present the Voxel-informed Language Grounder (VLG), a language grounding model that leverages 3D geometric information in the form of voxel maps derived from the visual input using a volumetric reconstruction model. We show that VLG significantly improves grounding accuracy on SNARE, an object reference game task.At the time of writing, VLG holds the top place (anonymized) on the SNARE leaderboard, achieving SOTA results with a 1.9% absolute improvement on grounding geometric descriptions and 1.7% overall improvement on all descriptions.
0 Replies
Loading