3D Scene Completion

3D scene completion reconstructs parts of an environment hidden or missing from partial observations, supporting tasks such as navigation, planning, and novel-view rendering. Systems use voxel and bird’s-eye-view networks, point-cloud or continuous geometry representations, and Gaussian splatting to predict shape and semantic occupancy. Transformer and diffusion models, often guided by visibility, depth, or semantic priors, generate plausible structure in unobserved regions; evaluation includes urban LiDAR benchmarks such as SemanticKITTI and indoor datasets such as NYUv2 and ScanNet. A central challenge is filling large gaps without hallucinating geometry or appearance that conflicts with observations, while remaining reliable across sensors, environments, and real-time constraints.