| With the growing proliferation of cameras and sensors, video and multimedia are becoming an indispensable part of our daily lives. The unprecedented production and consumption of visual information are making autonomous video analysis one of the most active and challenging research areas.; In this work I will focus on building a bridge between video data understanding and machine learning methods, which is a promising, but still not fully explored direction. I will specifically focus on model-based video analysis utilizing statistical graphical models. The core of our work is the investigation and design of clustering, inference, and learning algorithms applied to the video data. I will show how the inference in these models can be used to answer a number of queries useful for video analysis and processing.; Having the applications as the ultimate goal, I will demonstrate the algorithmic techniques for speeding up the naive learning in the graphical models by orders of magnitude. In that sense, I will investigate signal processing techniques, approximate methods, and online learning. I will demonstrate how the theory and algorithms usefully apply to the variety of tasks ranging from video clustering and stabilization, to video retrieval and building of the similarity measures between the distributions. |
