007 : Dense 3D reconstruction using moving monocular camera (2019)

A set of frames captured using moving Raspberry Pi3 in textured environment between rough centimeter offsets. Using initial guess poses, fundamental matrix is estimated, and a set of corresponding points are generated between reference image and current image based on geometrically constrained ORB matches. Sparse 3D point set is then generated using triangulation and full system is bundle adjusted iteratively into optimal configuration. Then dense matching is performed between the views. This step requires rectification using 3D poses to get horizontally aligned images. As last step, the resulting dense point clouds and depth maps are illustrated.

Input image and dense point cloud using the algorithm.

RPi3 setup with example reference image.

Stereo baselines to reference illustrated.

Dense depth maps illustrated.

Dense depth maps illustrated with overlaid textures.