Fitting Flashcards
Generalized hough transform
Used to find arbitrary shapes
Visual word
A target shape that wants to be detected
To train to find an arbitrary shape you
Create a code book of spaces nearby the targeted shape
For each interest point store all possible positions in code book
Testing to find arbitrary shape
Extract and match patches to code book entry
Cast votes for possible positions of objects center
Searching for maxima I’m voting space
Extract weighted segmentation mask
If we know which points belong to the line how do we find optimal parameters?
Least square
What if there are outliers to a line?
We use robust fitting
What if there are too many lines?
Voting methods: hough transform
What if we are not sure if it’s even a line?
Model selection (not covered)
Least square for criteria
The measure of how well a model fits a set of n observations
The best measure has the model minimizing its criteria
Problem with vertical least squares
Not rotation invariant
Fails completely for vertical lines
Random sample consensus
RANSAC
Very general framework for model fitting in the presence of outliers
RANSAC outline
Choose a small subset of points uniformly at random
For a model to that subset
Find all remaining points that are close to the model and reject the rest as outliers
Do this many times and choose the best model
Choosing RANSAC parameters
N = initial number of points
K= number of samples
Choose k that one sample is free from outliers
T= distance threshold
D= consensus set size
RANSAC pros
Simple and general
Applicable to many different problems
Often works well on practice
RANSAC CONS
Lots of parameters to tune
Doesn’t work well for low online ratios
Ca by always get a good initialization of the model based on min number of samples
Incremental line fitting
Puts all points on curve list in order along the curve
Empty the line point list
Empty the line list
Until there are too many points on the curve
For points to curve, check if good enough
K-means line fitting
Hypothesize k lines
Until convergence
Allocate each point to the closer line
Refit lines
