NeRF for Outdoor Scene Relighting

AbstractPhotorealistic editing of outdoor scenes from photographs requires a profound understanding the image formation process and an accurate estimation scene geometry, reflectance illumination. A delicate manipulation lighting can then be performed while keeping albedo geometry unaltered. We present NeRF-OSR, i.e., first approach for relighting based on neural radiance fields. In contrast to prior art, our method allows simultaneous illumination camera viewpoint using only collection photos shot in uncontrolled settings. Moreover, it enables direct control over illumination, as defined through spherical harmonics model. For evaluation, we collect new benchmark dataset several sites photographed multiple viewpoints at different times. each time, \(360^\circ \) environment map is captured together with colour-calibration chequerboard allow numerical evaluations real data against ground truth. Comparisons SoTA show that NeRF-OSR controllable higher quality realistic self-shadowing reproduction. (see project web page https://4dqv.mpi-inf.mpg.de/NeRF-OSR/)