C++ implementation (in ROS) of continuous fusion algorithm from Deep Continuous Fusion for Multi-Sensor 3D Object Detection. I can share a fast, vectorized SciPy version if there is any interest (update: see issues page).
Comments on coordinate systems:
Points x_velo in velodyne coords are sent to points x_image
in the image plane of camera 2 using (7) in Geiger et al.:
x_image = P * R * T * x_velo,
where:
x_velo - point in velodyne coords (x, y, z, 1).
T - velodyne coords to unrectified cam0 coords.
R - unrectified cam0 coords to rectified cam 0 coords,
P - rectified cam0 coords to image plane of cam2,
x_image - point in image of cam2 (u, v, 1).
(We use left-multiplying active/alibi coordinate transformations.)
Diagram borrowed from paper (we skip MLP):
