We trained a deep learning model to densify a sparse point cloud of a 3-D scene given an RGB image of the scene.
Compared to other learning methods, ours has 80% fewer parameters while outperforming others thanks to the mesh triangulation and linear interpolation used as pre-processing steps.
The paper describing our method has been accepted by the International Conference on Robotics and Automation (ICRA), 2020.
Thanks for pointing that out. First time doing "open-source" (well it seems it's not really open-source according to the modern definition). I'd like to use a more permissive license, but it's up to UCLA.
The original D435 does not have an IMU. But the D435i version has an IMU. We use it for our other projects which require the dense depth. But the SLAM system itself should work with only RGB and IMU after some calibration and parameter tuning.
The auto-calibration simply finds the spatial alignment between the camera and the IMU. If bad data are present, one needs some outlier rejection mechanism to filter out them. Auto-calibration alone does not provide that ability.
ROS makes the inter-process communication much easier if the SLAM system is incorporated as one component of a much bigger system. But you don't have to use ROS for that. We actually provide the ability to run it without ROS. Also, with ROS, it's easier to communicate with sensors given that the sensor drivers have been wrapped into ROS nodes.
This is part of my research as a graduate student at UCLA Vision Lab. The SLAM system is Extended Kalman Filter (EKF) based, has features (landmarks) in the state, and jointly estimates the pose of the camera and the location of the landmarks. It runs at 140 Hz on a PC and is much faster than (some if not all) existing open-source VIO systems.