אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Fast and reliable bundle adjustment is essential in many structure from motion (SfM) related applications such as mobile vision, augmented reality and robotics. In this talk an incremental and computationally efficient method for bundle adjustment will be presented. The method incorporates two key ideas to substantially reduce the involved computational cost, compared to a conventional bundle adjustment. First, the cost function is formulated in terms of multi-view constraints instead of re-projection equations, thereby eliminating the 3D structure from the optimization. If required, the observed 3D points, or any part of them, can be reconstructed based on the optimized camera poses. A formulation with three-view constraints is presented. As opposed to applying only epipolar geometry constraints, this formulation allows a consistent motion estimation also when the camera centers are co-linear. The second component is incremental smoothing, recently developed in the SLAM community, which adaptively identifies the variables that need to be recomputed at each step. The optimization problem is formulated in terms of a factor graph, and it is shown how to perform an efficient inference that typically involves only a fraction of the camera poses. In the final part of the talk, the application of the proposed method to multi-robot localization will be outlined and insights from a graph-based filtering approach for solving the same problem will be provided.