Creating high quality three-dimensional models of indoor scenes promises many applications in augmented reality, robotics, and object scanning. One popular approach for 3D model reconstruction is dense visual Simultaneous Localization and Mapping (SLAM). Existing dense SLAM methods provide accurate model reconstruction on a small-scale (using model-based techniques) or accurate camera trajectory estimation on a large-scale (using pose-graphs), but not both. Deformation graphs offer the best of both worlds as they can be applied to large-scale environments while directly optimizing model reconstruction, theoretically resulting in higher model accuracy. We will investigate how much deformation graphs embedded in the model’s surface will improve large-scale scene reconstruction quality

I worked on this project this summer as part of the Perception Systems Group at Draper Laboratory, modifying the iterative closest point component of the KinectFusion algorithm to make scene reconstruction more robust to user data collection approaches. I hope to gain a deeper understanding of how SLAM works and am excited to see how much deformation graphs can improve the current project’s performance.