3D Occupancy Prediction
Overview
The field predicts which parts of 3D space are occupied, what they contain, and sometimes how they will move, supporting navigation and autonomous driving. Common approaches lift camera features into voxel or bird’s-eye-view grids, fuse LiDAR or radar, or use sparse and continuous representations such as 3D Gaussians; training can rely on dense labels, depth cues, or pseudo-labels from pretrained models. Current work extends single-frame semantic prediction toward temporal forecasting, panoptic tracking, open-vocabulary understanding, and online maps for robots in indoor, outdoor, and specialized settings. A central tension is balancing dense, accurate geometry with efficient inference and reliable generalization, while ensuring that better occupancy scores translate into safer downstream planning.
Latest papers
113 papers21 papers in the last four weeks, up 24% on the four weeks before. 0.1% of all new papers.