How to measure the repeatability and correspondence of a feature detector?
I have two images of the same model from different viewpoints. The images are partially overlapped. I want to compare the performance of feature detectors like Harris Corner Detector, ORB, KAZE and MSER in OpenCV Python. For performance comparison, I have decided to follow a paper where the comparison was done by evaluating the repeatability and correspondence. Since I am just a beginner to stuff like this, I am having trouble understanding how exactly to measure the repeatability and correspondence.
The paragraph below is taken from a research paper I have been reading. I want to follow it’s procedure.
The objective of this experiment is to measure the repeatability and accuracy of the detectors: to what extent do the detected regions overlap exactly the same scene area (i.e., are the pre-images identical)? How often are regions detected in one image without the corresponding region being detected in another? Quantitative results are obtained by making these questions precise. The ground truth in all cases is provided by mapping the regions detected on the images in a set to a reference image using homographies. The basic measure of accuracy and repeatability we use is the relative amount of overlap between the detected region in the reference image and the region detected in the other image, projected onto the reference image using the homography relating the images.
My questions are:
-
What is the difference between the repeatability and correspondence? How do I calculate them?
-
"The ground truth in all cases is provided by mapping the regions detected on the images in a set to a reference image using homographies." Why do we need the ground truth and homography? How exactly should I calculate the ground truth and homography? Is it manually done or is there some algorithm?
Any help will be deeply appreciated.
