Improved mechanisms for three-dimensional image reconstruction are disclosed. A user is presented with a scene image and allowed to select elements of the image, such as planes, within which features are to be detected. The features are detected and tracked and objects (such as planes) are constructed. The user is allowed to revise the constructed objects and may be allowed to repeat image element selection, with feature detection and tracking and object construction being repeated. When element selection and object construction and user revision are completed, a three-dimensional reconstruction of the scene image is computed.