Intro
Direct vs Indirect(SLAM)
센서값을 그대로 쓰냐 vs sensor input을 시스템에 넣기 전에 해석을 거쳐서 input을 넣느냐
Dense vs Sparce
대부분의 pixel을 쓰냐 vs 특정한 pixel만 쓰냐
Main
Direct sparse model
- Calibration
sensor resonpe function을 만들어서 사용한다.
- Model Formulation
residual pattern을 만든다. Expectation을 구하는데, time과 affine brightness transfer function을 사용한다.
Visual Odometry Front-End
keyframe management
1. Initial frame tracking - projection active points
2. Keyframe creation - many keyframes as many as possible
3. Keyfrme Marginalization - removing the keyframes in ceratin rules
candidate management
1. Candidate point selection
2. Candidate point traking - point are tracked in the next frame
3. Candidate point activation - fill new points when old points are deleted
4. Outlier and occlusion detection - discard points when the error is too large
Result
TUM mono-VO dataset
DSO라는 방법이 error에 대해서 ,,,,,