Learning to select long-track features for structure-from-motion and visual SLAM

Abstract

With the emergence of augmented reality platforms, Structure-From-Motion or visual SLAM approaches have regained in importance in order to deliver the next generation of immersive 3D experiences. As a new quality is achieved by deployment on mobile devices, computational efficiency plays an important role. In this work, we aim to reduce complexity by limiting the number of features without sacrificing quality. We select a subset of image features, using a learning based approach. A random forest is trained to pick 2D image features which are likely to be significant for a 3D reconstruction. Additionally, we aim for an objective that selects long track features, so that they can be “re-used” in multiple frames. We evaluate our feature selection technique on real world sequences and show a significant reduction of image features and the resulting decreased computation time is not effecting the accuracy of the 3D reconstruction.