0. Introduction

• Automatic extractoin of “meaningful” information from images and videos

• semantic informatino

  • which class belongs parts to the image to?

• geometric information

  • 3D structure, horizon, etc.

• inverse problems

• computer vision:

  • take images and create model

• computer graphics:

  • take model and create images

• field studying relationship between

  • cameras

  • 3D struture

  • image features

• when there are feature correspondences between images taken from varying viewpoints

• -> image features usually points or lines

Camera position:

• absolute position of camera

camera pose:

• rotation and translation

• -> between cameras…

camrea trajectory:

• line connecting different cameras…

• process of incrementally estimating the pose of the vehicle

• by examining the chagnes that motion induces on the images of its onboard cameras

• visual odometry + loop detecoin & closure

• -> VO affected by noise

• SLAM guarantees global consistency

• structure from motion

• more general than SLAM

• -> works from unordered image sets (i.e. internet-sourced images…)

1. image pair

2. feature detection

3. feature matching/tracking

4. camera pose estimation

5. 3D reconstructoin

6. local optimization

• we got the image correspondance

• we found out about the camera poses

• from these, we reconstructed a 3D point cloud

• => but might be unsatisfactory…

=> need to simultaneously optimize camera poses (i.e. absolute) and the point cloud itself…

=> so called bundle adjustment