Error Functions
Choose a parametric model to represent a set of features.
For example: line detection
Issues
- Noise int eh measured feature locations
- Extraneous data: clutter (outliers), multiple lines
- Missing data: occlusion
Vertical Distance
Standard least squares approximation. But the problem with vertical least-square error is that it causes huge discrepancies as
Total Least Squares
A line is represented as: Unit Normal: The error metric between a point and a line:
Find (a, b, d) to minimize the sum of squared perpendicular distances
where and is a matrix of differences of averages for each point.
We can solve this with the SVD trick.
Least Squares as likelihood maximization
Generative Model
Assume the line points are corrupted by Gaussian noise in the direction perpendicular to the line.
Robust Estimator
General approach: minimize
where : residual of point w.r.t. model parameter
: robust function with scale parameter
For example: For a small residual (i.e. small error) relative to , this behaves much like the squared distance: we get . As grows, we get
The error function we defined is very sensitive to scale.