Speed Up Light Field Synthesis from Stereo Images

In this paper, we focus on the speedup of a learning-based light field synthesis pipeline. The pipeline involves a disparity estimation neural network and a light field blending component. The former achieves high speed performance through the use of feature extraction and multi-stage disparity refinement, while the latter warps and merges coarse light fields generated from the left and right disparity maps in a novel and efficient way. The pipeline can produce a full light field in less than 1/10 of a second, while retaining fairly reasonable image quality. The model itself has a very low parameter count, which is ideal for devices with limited computational power.