Generative Semantic Scene Completion
The paper turns LiDAR scene completion into a discrete-diffusion generation and refinement problem.
It introduces PS³ to generate paired sparse LiDAR scans and dense semantic completions, then trains with that corpus alongside SemanticKITTI. Its SGSC model completes scenes from noise, conditioned on a sparse scan through BEV semantic and 3D feature streams. A related S²D² step can refine SGSC or other SSC outputs without retraining the base model. The authors report 38.8% mIoU on the SemanticKITTI hidden test under causal, single-sweep, single-sample conditions. HF Daily Papers' note
It introduces PS³ to generate paired sparse LiDAR scans and dense semantic completions, then trains with that corpus alongside SemanticKITTI. Its SGSC model completes scenes from noise, conditioned on a sparse scan through BEV semantic and 3D feature streams. A related S²D² step can refine SGSC or other SSC outputs without retraining the base model. The authors report 38.8% mIoU on the SemanticKITTI hidden test under causal, single-sweep, single-sample conditions. HF Daily Papers' note
score 4