showing results for example stereo record 0507_12-39-44

Left Image
left image
Right Image
right image. translation = [0.483, 0.029, 0.066], axis rotation 1.05 degree
'Right' Image
distance
A correlation based shift applied to the right image as rectfication. Disparity calculation is performed with the RAFT stereo alogirthm (see RAFT paper of Lipson, Teed and Deng ) implementation from github.com/princeton-vl/RAFT-Stereo . The distance in meter is calculated with the parameters of a Samsung S7 camera (FOV 41 degree.). The distance map is masked for the pixels of the building objects.
'Right' Image
semantic segmentation of the image with the keras-segmentation library
For the segmentation the moidel trained for the ADE20K was used. for the labels see ADE20k labels.

The resulting depth map converted to point a 3D cloud: