image matching 2 Flashcards
(17 cards)
What is the goal of matching image points?
To find corresponding points between an image pair, enabling tasks like alignment and 3D reconstruction.
What are two main ways to obtain corresponding image points?
Manually (by an operator) and automatically (by an algorithm).
What is cross-correlation (CC) used for in image matching?
To find where a template appears in a larger image by comparing similarity across all possible positions.
What are key assumptions of cross-correlation?
Images only differ by translation, brightness, and contrast.
What is the best estimate of offset in cross-correlation?
The offset u that maximises the cross-correlation coefficient.
What is the role of normalised cross-correlation?
To provide a more robust similarity measure that accounts for brightness and contrast differences.
What is exhaustive search in template matching?
Computing similarity for all possible offsets and selecting the one with the highest score.
What is the coarse-to-fine search strategy?
A method using image pyramids to search for template matches at multiple scales, improving efficiency.
What is the main limitation of cross-correlation?
It performs poorly with occlusions, large rotations (>20°), and large scale changes (>30%).
What is least squares matching (LSM)?
A precise method for fitting a model to data by minimising the squared residuals between observed and predicted values.
What type of models can be used in LSM?
Parametric models such as lines, circles, or transformations (e.g., affine).
What are the two broad approaches to image alignment mentioned in the slides?
Direct (pixel-based) alignment and feature-based alignment.
What does direct alignment aim to do?
Align images such that the pixel values match as closely as possible.
What transformation types are mentioned in the slides?
Similarity, Affine, and Projective (Homography) transformations.
What kind of transformation is suitable for roughly planar scenes?
Affine transformation.
How many point correspondences are required to solve for an affine transformation?
At least three, as each provides two independent equations.
What is the purpose of fitting a transformation in image alignment?
To find the model parameters that best map matched features from one image to another.
