Realistic 3D reconstruction from an image sequence

A factorization-based 3D reconstruction system is implemented to recover 3D models from an image sequence. The image sequence is obtained from uncalibrated perspective cameras. First, the KLT feature tracking method is used to collect many matching feature points for all images. Then, these data are provided to the 3D reconstruction module to get the projective reconstruction. The projective reconstruction is converted to a Euclidean one by enforcing metric constraints. Finally, realistic 3D models can be obtained by texture mapping. To increase the portability of the developed system, the whole system is implemented in C++ language. Qt library is used to implement the interactive system user interface and OpenGL graphics library is used to implement the texture mapping routine. The model visualization program written in MFC can display both wire-frame objects and texture-mapped objects. Several experimental results on both synthetic and real image data are included to demonstrate the effectiveness of the developed system. The developed system will be extended to integrate several data sets from multiple views to make a complete reconstruction of the scene.