Finding correspondences
A lot of computer vision problems can be dealt with finding the correspondences between two images, The process can be described as follows
flowchart TD
A[DETECTION]
B[DESCRIPTION]
C[MATCHING]
A --> B
B --> C
Where:
- in detection phase interesting point of an image are detected (keypoints)
- in description phase for each keypoint a descriptor based on the neighborhood is computed
- in the matching phase descriptors of keypoints detected from a reference image are compared against descriptors of keypoint from different images
The algorithms for the detection and description phase must have the following properties
| DETECTORS | DESCRIPTORS |
|---|---|
| REPEATABILITY it should find the same keypoint in the image despite the transformations undergone by the images |
DISTINCTIVENESS/ROBUSTNESS TRADE OFF it should capture the salient information in the neighborhood of a keypoint and also avoid noise effects from change of light intensities |
| INTERESTINGNESS It should find points with informative surroundings as to enable the matching process |
COMPACTNESS description should be concise as possible in order to improve the matching process |
Keypoints: choose the best candidate
keypoints are points that contains the most information in an image, edges are bad candidates as along the perpendicular direction they are pretty similar and cannot be distinguished, points that show high variance in all directions are best suited for the purpose such as corners
