Real-time Neural Rendering of LiDAR Point Clouds
Rendering high-fidelity color images directly from static LiDAR scan point clouds remains challenging due to severe artifacts. This paper proposes a real-time neural rendering method that requires neither scene-specific training nor heavy preprocessing. Our approach introduces a synergistic framework combining a depth-guided U-Net and a depth-aware heuristic pre-filter: the U-Net performs end-to-end mapping via depth-guided point cloud projection, while the pre-filter suppresses artifacts without requiring registration ground truth. Leveraging synthetic data augmentation and GPU-optimized inference, the method achieves >30 FPS rendering on commodity GPUs. To our knowledge, this is the first neural point cloud renderer achieving robustness, high fidelity, and generalizability without any ground-truth annotations. Quantitative and qualitative evaluations demonstrate superior rendering quality and speed over state-of-the-art methods.