Half-occluded regions: The key to detecting a diverse array of defects in S3D imagery

摘要

Half-occlusions, the regions of a binocular image which are only visible in one view, convey important depth ordering information that is left aside by most stereo computer vision algorithms and stereo post-production practices of the stereoscopic 3D (S3D) film industry. We demonstrate the efficacy of using rules governing the geometry of half-occlusions to detect occlusion-stereopsis conflicts in S3D imagery. Experimentation shows that our half-occlusion based technique allow high accuracy detection of global and local pseudoscopy, stereo window violations and rotoscopy mismatches at a low computational cost for both stereo matched and synthetic disparity maps. Such a detection algorithm could prove useful in quality control of S3D imagery for cinema and virtual reality.