Effective Multi-sensor Conditioning for Street-view Novel-view Synthesis
Researchers have developed StreetNVS, a novel video diffusion framework designed for synthesizing new views of driving scenes. This method effectively fuses data from multiple sensors, including LiDAR, cameras, and ego-motion, to generate high-quality novel views. StreetNVS significantly outperforms existing methods, even when using substantially sparser LiDAR data, and demonstrates capabilities in generating views along extreme out-of-trajectory paths. AI
IMPACT Enhances generative capabilities for scene reconstruction and simulation in autonomous driving contexts.