feature_based_matching_flashcards
(20 cards)
What is a key challenge in feature-based matching?
Finding correspondences when they are not known in advance.
What is a keypoint in image processing?
An image location where a feature descriptor is computed.
What does a feature descriptor do?
Summarises the local image structure around a keypoint.
What are the advantages of local features?
Locality, distinctiveness, quantity, efficiency, and extensibility.
What is the Förstner-Operator used for?
Identifying locally distinct points that can be well localised.
What is the purpose of the Harris Detector?
To find corners by measuring intensity changes in all directions.
What do the eigenvalues λ1 and λ2 represent in the Harris Detector?
They represent intensity changes in orthogonal directions; both large indicates a corner.
What is the corner response function R?
A computed value that indicates how ‘corner-like’ a pixel is.
How does the Harris Detector select final keypoints?
By thresholding R and taking local maxima.
What is the simplest feature descriptor?
A 3x3 window around the keypoint.
What are examples of advanced descriptors?
SIFT, SURF, HOG, GLOH.
How does the SIFT descriptor work?
It uses gradients in a 16x16 region to build 128D orientation histograms.
What is the basic idea of feature matching?
Compare descriptors across images and find the one with minimum distance.
What is SSD in feature matching?
Sum of squared differences between descriptor entries.
Why is choosing a good threshold important in feature matching?
To maximise true positives and minimise false positives.
What is the correspondence problem?
Incorrect matches due to ambiguous descriptors or noise.
What is the goal of RANSAC?
To robustly fit a model by separating inliers from outliers.
What are the steps of the RANSAC algorithm?
Sample data, compute model, count inliers, repeat until best model is found.
What does feature-based alignment involve?
Extracting features, computing matches, hypothesising and verifying transformations.
Why is RANSAC essential in feature-based matching?
It provides a way to estimate models despite many incorrect matches.
