Regularized Adaboost for RGBD video content identification

Content identification;fingerprinting;learning theory;mutual information;Kinect camera;depth video

Yu, Honghai ; Moulin ; Pierre

关键词: Content identification;    fingerprinting;    learning theory;    mutual information;    Kinect camera;    depth video;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/42242/Honghai_Yu.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This thesis presents three contributions. First, we provide an information theoretic analysis to a recently developed learning-based content identification (ID) algorithm, symmetric pairwise boosting (SPB). Second, we propose a regularized Adaboost algorithm, which tackles SPB’s implicit assumption that video segments are statistically independent. Finally, we develop the first hybrid content ID system for synchronized RGB and depth (RGBD) videos. Experimental results show the regularized Adaboost algorithm vastly outperforms SPB for all considered distortions, while the hybrid system further improves the content ID performance of regularized Adaboost relative to RGB-alone or depth-alone content ID systems.