3D Scene Understanding

3D scene understanding studies how to recover the geometry, objects, relationships, and often motion of environments from images, depth, and LiDAR. Common approaches include point-cloud and voxel networks, occupancy prediction, Gaussian splatting, and 3D scene graphs, increasingly combined with vision-language models for grounding and reasoning. These representations support tasks such as robot navigation and manipulation, autonomous driving, scene reconstruction, and open-vocabulary object retrieval. A central challenge is making systems both spatially reliable and efficient in large, changing scenes, while benchmarks continue to expose gaps in metric reasoning, cross-view consistency, and generalization.