ICCV Daily 2019 - Thursday

Amnon Geifman is an MSc research assistant in Ronen Basri’s lab at Weizmann Institute in Rehovot, Israel. He speaks to us ahead of the team’s poster session today. Amnon tells us their work is about essential matrix averaging , which is the extraction of the 3D structure and camera motion from a raw collection of essential matrices . This is done by defining a necessary and sufficient condition on the multi-view essential matrix. He explains further: “What we have done is to build a block matrix, called a multi-view essential matrix, which is a block matrix that contains in the i'th j'th block the essential matrix between the ith frame and jth frame. We then prove the necessary and sufficient conditions for such a matrix to be consistent with n camera matrices. This means if a multi-view matrix was generated from n camera matrices, it needs to satisfy those conditions and vice versa. Then using these conditions, we formalize a rank-constrained optimization to recover a consistent multi-view essential matrix from a noisy one. After the optimization, we also formalize a method for extraction of camera parameters from a consistent multi- view essential matrix.” The idea for the work came from previous work that the team had published on fundamental matrix averaging at CVPR 2019 . This time they chose to do essential matrix averaging because it’s more practical and people use it more often in the industry, so it has many more applications than the fundamental matrix. Amnon says most of the existing methods are doing rotation averaging , and then by using the rotations, doing translation averaging. The rotation and translation are extracted from 10 Poster Presentation DA I L Y Algebraic Characterization of Essential Matrices and Their Averaging in Multiview Settings "Means if a multi-view matrix was generated from n camera matrices, it needs to satisfy those conditions and vice versa."