Dynamic Novel View Synthesis
Overview
Dynamic novel view synthesis studies how to render a scene from new camera positions while preserving moving objects, changing appearance, and temporal continuity. Methods include neural radiance fields and 4D Gaussian splatting, which represent scene geometry and motion, as well as video diffusion models guided by camera paths, depth, optical flow, or multi-view observations. Researchers evaluate these approaches on synthetic dynamic scenes, captured human and object videos, and driving datasets such as Waymo, with tests ranging from nearby views to large viewpoint changes and future-frame synthesis. A central challenge is balancing realistic detail and fast rendering with geometric and temporal consistency, especially when training footage is sparse, monocular, or captured from cameras with little overlap.
Latest papers
41 papers4 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.