Changqing Zhou

Achieve state-of-the-art open-vocabulary occupancy prediction in complex indoor scenes by using only binary occupancy labels and a novel Gaussian-based representation.

By extending surface points inward along camera rays and representing volumetric samples as Gaussian primitives, GPOcc achieves state-of-the-art results in monocular and streaming occupancy prediction, outperforming existing methods by a significant margin while also running faster.