Machine Learning for Robotic Exploration
Strojové učení pro robotickou exploraci
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
Tato práce navrhuje, implementuje a vyhodnocuje nový způsob trénování sítí pro doplňování hloubky pomocí chyby rekonstrukce mapy. Byl také implementován tradiční způsob trénování hloubkových doplňovacích sítí pomocí MSE a obě metody byly porovnány. Do učebního procesu jsme zahrnuli modul diferencovatelného hustého SLAMu a model jsme vyhodnotili na datové sadě KITTI. Implementovali jsme síť pro doplňování hloubky, aby byl vstup do SLAMu hustší a poskytl tak více korespondencí, se kterými lze pracovat. Pomocí sítě pro doplňování hloubky jsme byli schopni získat hustší hloubkové mapy. Trénování s chybou rekonstrukce mapy přineslo podobné výsledky jako tradiční metody. Hustší data nezvýšila přesnost lokalizace SLAM, a to hlavně proto, že model vytvořil příliš mnoho odlehlých hodnot a ztížil práci pro SLAM.
This work, we proposes, implements and evaluates a novel way to train depth completion networks using map reconstruction error. A traditional way to train depth completion networks with MSE was also implemented, and both methods were compared. We have included a differentiable dense SLAM module in our learning pipeline and evaluated the model on the KITTI dataset. We introduced depth completion network to make input to SLAM more dense in order to provide more correspondences to work with. Using the depth completion network, we were able to obtain denser depth maps. Training with the map reconstruction error yielded results similar to those of traditional methods. Denser data did not increase SLAM localization accuracy, this was mainly because the model introduced too many outliers and made it difficult for SLAM to work.
This work, we proposes, implements and evaluates a novel way to train depth completion networks using map reconstruction error. A traditional way to train depth completion networks with MSE was also implemented, and both methods were compared. We have included a differentiable dense SLAM module in our learning pipeline and evaluated the model on the KITTI dataset. We introduced depth completion network to make input to SLAM more dense in order to provide more correspondences to work with. Using the depth completion network, we were able to obtain denser depth maps. Training with the map reconstruction error yielded results similar to those of traditional methods. Denser data did not increase SLAM localization accuracy, this was mainly because the model introduced too many outliers and made it difficult for SLAM to work.