Adaptive 3D Video Parallax Generation Based on Neural Radiation Field

Stereo images generated by existing autostereoscopic transformation methods usually have many problems, such as distortion of the occluded region, poor visual comfort, etc. Therefore, we propose a Neural Radiation Field based framework for adaptive generation of parallax for the 3D video to achieve steerable high-quality 3D video. Firstly, we introduced the NeRF with bidirectional flow in the time domain to generate the dynamic video field with a new perspective, thus that could accurately and efficiently predict image information located in occluded areas. Secondly, an adaptive and accurate parallax control method was proposed, which can adaptively calculate ideal parallax according to the viewing conditions and the video scene characteristics, to generate stereoscopic video with remarkable stereoscopic effects and comfortable vision. Experimental results show that our method could generate high-quality 3D videos with a lower visual fatigue index and 3D sense.