A locally optimized scene reconstruction from three uncalibrated color images

This paper presents a new robust scene-recognition algorithm using points and lines in three uncalibrated color images under natural or artificial lighting without knowing the poses of the camera. Local optimization is employed by the algorithm in finding point matches to guarantee accuracy in the case when object surface is extremely uneven of having abrupt changes, thus causing a wide range of point movements between consecutive images. The algorithm is iterated to obtain maximum features until no improvement can be achieved. Experiments using a large number of images show excellent performance by this new algorithm.