2020/10/19 by Juan J. Gómez Rodríguez, José Lamarca, Rodríguez, Juan J. Gómez +7 · 2 citations
Computer Science · Earth and Planetary Sciences · Engineering · #3D Surveying and Cultural Heritage #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #I.4.5 #I.4.6 #I.4.8 #Multimodal Machine Learning Applications #Robotics and Sensor-Based Localization
paper · pdf · doi:10.48550/arxiv.2010.09409
openalex publication_date 2020/10/19 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Conventional SLAM techniques strongly rely on scene rigidity to solve data\nassociation, ignoring dynamic parts of the scene. In this work we present\nSemi-Direct DefSLAM (SD-DefSLAM), a novel monocular deformable SLAM method able\nto map highly deforming environments, built on top of DefSLAM. To robustly\nsolve data association in challenging deforming scenes, SD-DefSLAM combines\ndirect and indirect methods: an enhanced illumination-invariant Lucas-Kanade\ntracker for data association, geometric Bundle Adjustment for pose and\ndeformable map estimation, and bag-of-words based on feature descriptors for\ncamera relocation. Dynamic objects are detected and segmented-out using a CNN\ntrained for the specific application domain. We thoroughly evaluate our system\nin two public datasets. The mandala dataset is a SLAM benchmark with\nincreasingly aggressive deformations. The Hamlyn dataset contains\nintracorporeal sequences that pose serious real-life challenges beyond\ndeformation like weak texture, specular reflections, surgical tools and\nocclusions. Our results show that SD-DefSLAM outperforms DefSLAM in point\ntracking, reconstruction accuracy and scale drift thanks to the improvement in\nall the data association steps, being the first system able to robustly perform\nSLAM inside the human body.\n