Novel view synthesis under trajectory changes is a key challenge for autonomous driving simulation. According to a new arXiv paper, current methods have difficulty producing consistent videos when asked to extrapolate to viewpoints not seen during training, such as lane shifts.
The authors introduce GeoV2V, a geometry-grounded video diffusion model. The approach incorporates geometric information to ground the generation process, aiming to improve consistency for unseen viewpoints.
The abstract does not provide quantitative results or comparisons, so the effectiveness of GeoV2V is not yet established from the announcement alone. The full paper is available on arXiv.