Skip to content

Implement REPROJ_AVG_DEPTH

Mateo de Mayo requested to merge mateosss/dynviof into xrtslam

This is a heuristic that replaces the old view_offset properties. As Basalt uses optical flow for feature matching, which in turn is a non linear optimization, a good enough initial estimate needs to be provided. Before the same pixel was being used (now called SAME_PIXEL), now we assume features in the left camera have a depth that is equal to the average of all currently known depths, and reproject that 3D point into the right camera (REPROJ_AVG_DEPTH). In general, it is a better guess as can be seen by enabling the UI toggles and enables cameras which are not almost parallel (like headset cameras) to work better (although it also improved significantly the features in my D455 which has pretty parallel cameras).

Depends on basalt-headers!3 (merged)

Edited by Mateo de Mayo

Merge request reports