Towards Real-Time and Adaptable LiDAR Scene Completion
RapidLiDAR claims full-scene LiDAR completion in 0.1 seconds, matching a typical 10 Hz automotive sensor rate.
The paper says the method learns its coarse scene initialization instead of relying on fixed noise or slower generative refinement. It predicts spatially varying displacements for partial input points, then refines them with multi-scale voxel and BEV features. On SemanticKITTI and KITTI-360, the authors report state-of-the-art-level completion while running 2.3 times faster than the fastest prior method. Accepted at ECCVW 2026, with code listed as available. HF Daily Papers' note
The paper says the method learns its coarse scene initialization instead of relying on fixed noise or slower generative refinement. It predicts spatially varying displacements for partial input points, then refines them with multi-scale voxel and BEV features. On SemanticKITTI and KITTI-360, the authors report state-of-the-art-level completion while running 2.3 times faster than the fastest prior method. Accepted at ECCVW 2026, with code listed as available. HF Daily Papers' note
score 4