Title: Not All Points Are Equal: Uncertainty-Aware 4D LiDAR Scene Synthesis

Abstract: The ability to construct accurate 4D environments from LiDAR sequences is essential for embodied AI; however, existing generative models typically allocate equal modeling resources to all spatial areas. This approach overlooks the significant variation in perceptual difficulty within a single scan, where distant surfaces, occluded edges, and small objects exhibit much higher uncertainty compared to clearly observed structures. To address this, we introduce U4D, a novel framework that utilizes spatial uncertainty to direct LiDAR scene generation through a "hard-to-easy" scheduling strategy. U4D generates per-point uncertainty maps based on Shannon Entropy calculated from a pretrained segmentor. It first employs an unconditional diffusion process to reconstruct high-entropy regions with accurate geometry, and then uses these structures as priors in a conditional completion stage to fill in the remaining areas. Additionally, a Mixture of Spatio-Temporal (MoST) block ensures cross-frame coherence by dynamically balancing spatial detail with temporal continuity. Comprehensive evaluations on the nuScenes and SemanticKITTI datasets show that our method achieves state-of-the-art results in terms of scene fidelity, temporal consistency, and downstream task performance.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC