Sparse 3D Reconstruction

This video shows a sparse 3D reconstruction of a small area of an indoor environment The camera was mounted on a cart and translated and rotated around the scene. 3D features are tracked over time with KLT tracker. Kalman filters estimate the optimal 3D position of the features over time. The ego-motion of the camera is computed minimizing the reprojection errors of triangulated 3D points in consecutive frames. Observe that, even though the ego-motion drifts over time, the overall estimation is good enough to allow a coherent 3D reconstruction of the scene. The colors of the features encode the depth (red is close, yellow is middle distance, and green is far). The gray scale value of the 3D points correspond to the original luminance value obtained from the image. The colored 3D points in the 3D reconstruction show the positions of the camera throughout the sequence. More details of our method is found here: H. Badino and T. Kanade: "A Head-Wearable Short-Baseline Stereo System for the Simultaneous Estimation of Structure and Motion". IAPR Conference on Machine Vision Applications (MVA), Nara, Japan, June 2011. Hernan Badino Carnegie Mellon University http://www.lelaps.de/