Multimodal indexing of presentation videos

Abstract

This thesis presents four novel methods to help users efficiently and effectively retrieve information from unstructured and unsourced multimedia sources, particular the increasing amount and variety of presentation videos such as those e-learning, conference recordings, corporate talks, and student presentations. We demonstrate a system to summarize, index and cross-reference such videos, and measure the quality of the produced indexes as perceived by the end users. We introduce four major semantic indexing cues: text, speaker faces, graphics, and mosaics, going beyond standard tag based searches and simple video playbacks. This work aims at recognizing visual content in the wild, where the system cannot rely on any additional information besides the video itself. For text, within a scene text detection and recognition framework, we present a novel locally optimal adaptive binarization algorithm, implemented with integral histograms. It determines of an optimal threshold that maximizes the between-classes variance within a subwindow, with computational complexity independent from the size of the window itself. We obtain character recognition rates of 74%, as validated against ground truth of 8 presentation videos spanning over 1 hour and 45 minutes, which almost doubles the baseline performance of an open source OCR engine. For speaker faces, we detect, track, match, and finally select a humanly preferred face icon per speaker, based on three quality measures: resolution, amount of skin, and pose. We register a 87% accordance (51 out of 58 speakers) between the face indexes automatically generated from three unstructured presentation videos of approximately 45 minutes each, and human preferences recorded through Mechanical Turk experiments. For diagrams, we locate graphics inside frames showing a projected slide, cluster them according to an on-line algorithm based on a combination of visual and temporal information, and select and color-correct their representatives to match human preferences recorded through Mechanical Turk experiments. We register 71% accuracy (57 out of 81 unique diagrams properly identified, selected and color-corrected) on three hours of videos containing five different presentations. For mosaics, we combine two existing suturing measures, to extend video images into in-the-world coordinate system. A set of frames to be registered into a mosaic are sampled according to the PTZ camera movement, which is computed through least square estimation starting from the luminance constancy assumption. A local features based stitching algorithm is then applied to estimate the homography among a set of video frames and median blending is used to render pixels overlapping regions of the mosaic. For two of these indexes, namely faces and diagrams, we present two novel MTurk-derived user data collections to determine viewer preferences, and show that they are matched selection by our methods. The net result work of this thesis allows users to search, inside a video collection as well as within a single video clip, for a segment of presentation by professor X on topic Y, containing graph Z.