| To record an exciting moment, we often capture scene activities from different viewpoints by moving the camera while capturing the video. During playback, it is desirable to have the ability to interactively control the timeline, e.g., for slow motion playback, and to control the viewpoint to view the activities.;In this thesis, we propose a probabilistic framework for reconstructing scene geometry and object motion utilizing prior knowledge of a class of scenes, for example, scenes captured by a camera mounted on a vehicle driving through city streets. In this framework, we assume the video camera is always calibrated, i.e., the intrinsic and extrinsic parameters are known all the time. We assume a single camera moving during capturing, but the framework can be generalized to multiple stationary or moving cameras as well. Traditional approaches try to match the points, lines or patches in multiple images to reconstruct scene geometry and object motion. The proposed framework also takes advantage of each patch's appearance and location to infer its orientation and motion direction using prior based on statistical learning from training data. The prior hence enhances the performance of geometry and motion reconstruction. We show that the prior-based 3D reconstruction outperforms traditional 3D reconstruction with synthetic data and real data for both stationary scenes and dynamic scenes, especially in the textureless areas for geometry estimation and faraway areas for motion estimation. |
