Blind Video Temporal Consistency

摘要

Extending image processing techniques to videos is a non-trivialrntask; applying processing independently to each video frame oftenrnleads to temporal inconsistencies, and explicitly encoding temporalrnconsistency requires algorithmic changes. We describe a morerngeneral approach to temporal consistency. We propose a gradientdomainrntechnique that is blind to the particular image processingrnalgorithm. Our technique takes a series of processed frames thatrnsuffers from flickering and generates a temporally-consistent videornsequence. The core of our solution is to infer the temporal regularityrnfrom the original unprocessed video, and use it as a temporalrnconsistency guide to stabilize the processed sequence. We formallyrncharacterize the frequency properties of our technique, and demonstrate,rnin practice, its ability to stabilize a wide range of popularrnimage processing techniques including enhancement and stylizationrnof color and tone, intrinsic images, and depth estimation.