Direct Linear Structure-from-Motion

Опубликовано 27 июня 2016, 20:58

Many structure-from-motion (SfM) systems take an incremental approach to reconstruct input images one by one. Global SfM methods that estimate all cameras simultaneously are highly desirable in terms of computation efficiency and result accuracy. We present a direct linear method for global SfM, which solve all cameras in a linear equation. We derive these linear equations from camera triplets, and generalize them to feature tracks to deal with weakly associated data. Experiments on various data demonstrate the superior efficiency and accuracy of our method.