FPSGen: Flexible Point Cloud Scene Generation with BEV-Supported Transport Flows

FPSGen introduces an innovative framework designed for versatile point cloud scene generation, tackling the discrepancies between training and inference found in current LiDAR-conditioned techniques. Conventional methods depend on partial scans for their initial setup, which results in issues like sparsity and visibility bias, causing incomplete geometry in occluded and distant areas. In contrast, FPSGen generates point sources independently of these partial scans by initially predicting a bird's-eye-view (BEV) prior that includes density, height, and mask channels from active cues. This density map is then utilized to create a BEV-supported point source, facilitating both unconditional and conditioned initialization. By eliminating reliance on LiDAR data, this framework can generate scenes from layout cues or alternative inputs. The teacher-student architecture with transport flows further enhances the refinement of the generated point clouds. Detailed in a paper on arXiv (2607.26645), FPSGen marks a notable leap forward in 3D scene generation for applications in autonomous driving and robotics.