RGBDTAM: A Cost-Effective and Accurate RGB-D Tracking and Mapping System
In this post, Alejo Concha Belenguer presents his new work RGBDTAM. This is a direct SLAM pipeline that runs in real-time and on a CPU. Code is available on github.
“The key ingredients of our approach are mainly two. Firstly, we evaluate different configurations for the error minimization in the 6DoF pose estimation. Specifically, these different configurations are related to dense vs semi-dense, photometric error vs geometric error and depth vs inverse depth error (in the geometric term). We show that the fusion of a semi-dense photometric and dense geometric (inverse depth) error for the pose optimization (Figure 1 in the paper) is the most accurate option. And secondly, we model the multi-view constraints and their errors in the mapping and pose tracking threads, which adds extra information over other approaches. We compare our system against ElasticFusion in the TUM dataset.”
Source code: https://github.com/alejocb/rgbdtam