PROJECT TITLE :

Parallax Tolerant Light Field Stitching for Hand-Held Plenoptic Cameras

ABSTRACT:

Hand-held plenoptic cameras may benefit from light field (LF) stitching as a means of expanding their field of vision (FOV). There are currently no LF stitching methods that are capable of accurately registering scenes with a substantial depth difference. New LF stitching methods are proposed in this research to address parallax in the LFs more correctly and flexibly. To begin, a depth layer map (DLM) is presented as a means of ensuring that each depth layer has an acceptable number of point feature values. Superpixel layer map (SLM) based on LF spatial correlation analysis is presented to refine the depth layer assignments in nondeterministic depth regions. The location dependent homography transformations are then precisely derived and LFs are warped to their appropriate position without parallax interference using DLM-SLM-based LF registration. The registration results are further fused using 4D graph-cut for improved LF spatial and angular continuity. The proposed method's superior performance in terms of subjective quality of the stitched LFs, epipolar plane image consistency in the stitched LF, and perspective-averaged correlation between the stitched LF and the input LFs is demonstrated for various scenes using horizontal, vertical, and multi-LF stitching.