C10 - Combination of Different Configurations

• alternatively estimate camera pose

• and (triangulate/reconstruct) 3D points

in monocular or stereo camera mode

1. pose initialization based on 2D-2D correspondance (-> find relative pose); left camera frame trated as global world frame

2. use the estimated camera pose and 2D-2D correspondances to triangulate initial 3D points

3. use 3D-2D correspondances to estimate camera pose of current frame (absolute -> i.e. w.r.t world)

4. use this absolute pose and new 2D-2D correspondances to triangulate new 3D points -> directly in world frame…

5. use new 3D-2D correspondances to further estimate pose of new frame

6. …

• goal: estimate relative pose between first camera (world frame) and second camera

• -> use 5-Point or 8-Point RANSAC

• then triangulate 3D point based on estimated pose

• scale ambiguity

• -> simply determine initial scale and continue on with it… (-> i.e. normalize translation vector)

• the larger the norm

• -> the larger the point cloud

• given a 3D point cloud and already associated 2D points (i.e. we have 3D-2D correspondances or already searched for them…)

• determine the absolute pose of the new view

• scale of extrinsic parameters aligned to pre-defined scale of reconstructed 3D points…

• use the estimated absolute camera pose to triangulate new 3D points in the world frame

• use 3D-2D correspondences to estimate the absolute pose of the fourth frame

1. initialization

2. absolute pose estimation from 3D-2D point correspondences

3. incremental 3D reconstruction an and absolute pose estimation

4. third step for next one…

• as we know depth -> need no triangulation

• use two images to create local map (-> based on disparity of dense 2D-2D correspondances)

• align this local map to the points in the world frame (i.e. an incomplete global map) -> based on 3D-3D correspondaences to estimate absolute camera pose (ICP)

• transform the reconstructed global map to the world frame -> to increment the global 3D map

• …

=> by this, we gradually extend our (point cloud) global 3D map

• stereo: can directly apply 3D-3D

• monocular: need pipeline of 2D-2D; 3D-2D; 3D-3D…

• => used in man-made environments (where we can apply our city models e.g. manhattan)

• -> perform camera pose estimation

• -> perform camera pose optimization

• same dominant directions should be observable in different images without any overlap (i.e. dominant directions do not suddenly change at some other place…)

• => this is a global constraint compared to feature correspondences

• -> as feature correspondences are only provide local constraints!

• use geometric contraints

• -> compute dominant directions from camera frame

• -> align these with a relative rotation between the camera frames

!!!! we do not consider estimating the rotation here!!!