Tobias Bertel, Yusuke Tomoto, Srinivas Rao, Rodrigo Ortiz-Cayon, Stefan Holzer and Christian Richardt

"Deferred Neural Rendering for View Extrapolation"

in ACM SIGGRAPH Asia 2020 posters

We capture an input video with a consumer camera, estimate camera poses, reconstruct a mesh and uv-map it.
We extend Deferred Neural Rendering [Thies et al. 2019] (blue) to enable smooth extrapolation of novel viewpoints (orange).

Image-based rendering methods that support visually pleasing specular surface reflections require accurate surface geometry and a large number of input images. Recent advances in neural scene 
representations show excellent visual quality while requiring only imperfect mesh proxies or no surface-based proxies at all. While providing state-of-the-art visual quality, the inference time of  learned models is usually too slow for interactive applications. While using a casually captured circular video sweep as input, we extend Deferred Neural Rendering to extrapolate smooth viewpoints  around specular objects like a car.

Submission video

Back to Top