Image Analysis II: 3D and Motion Reconstruction
The topic of this lecture is how to reconstruct a 3D geometry model of the world from a collection of images. We will first study the classical "structure-from-motion" pipeline, which determines camera pose from feature correspondences, and a 3D point cloud from feature triangulation. For this, we need solid knowledge on the principles of projection and the geometry of cameras. Later on, we upgrade the sparse point cloud reconstruction to dense models using e.g. dense stereo matching or RGB-D cameras like Kinect, and learn how to incorporate motion reconstruction for scene flow from multiple videos.
Topics in this course include:
- Image correspondence and alignment
- Geometry of perspective projection
- Camera calibration
- Two-view geometry and camera pose estimation
- Triangulation of point clouds
- Dense stereo matching and dense models
- Motion tracking and scene flow
Prerequisites are solid linear algebra, some calculus and statistics does not hurt. Image Analysis I is not required, but useful - we will make use of feature extraction methods, and again use MATLAB in the exercises. However, this can be picked up on the fly during the lecture.