The generation of realistic three-dimensional (3D) models of objects from video has been considered by the computer vision and graphics communities for several years. Several automatic methods have been proposed, but they suffer from ambiguities in image data, degeneracy of camera motion, and lack of appropriate feature points on the model surface. This paper proposes a system, called VideoTrace, to solve these problems by using user interaction and computer vision techniques.

The input video sequence is preprocessed to obtain a 3D point cloud using structure and motion analysis, and is also segmented such that a large number of small clusters of adjacent pixels have the same color. The 3D point cloud is overlaid on top of a frame from the video sequence, and a user can trace an object on any frame of the video. The inaccuracies in tracing are overcome by automatic refinement, and a polygonal face of the object is modeled by joining the line segments drawn by the user. Nonuniform rational B-splines are used to model the surface, and extrusions are appropriately handled. The system uses mirror planes to capture global properties of the surface, such as regularity and symmetry.

The scheme has been evaluated using a wide variety of video sequences, and the results are promising. VideoTrace combines the power of automatic and manual modeling systems. The system may be of help to users interested in accurate 3D reconstruction from video.