Feature Finding/Correspondence:
1)Detect Features – Find POI’s by a repeatable detector.
2)Find Corresponding Pairs by distinctive descriptors.
*In case of panoramic view – stitching – align images
The main issues in the process are:
1) Find most of the points of interest on both images, by a repeatable detector. (precision, repeatability)
2) Match corresponding points by a reliable and distinctive descriptor, a feature can be the matched to slightly similar feature but can’t be matched to a distinctively non similar feature that is very changed or different.(saliency/reliability/distinctiveness/matchability)
3)There is no need to have a feature everywhere (compactness/efficiency).
4)Locality -is not precision. The descriptor should be based on a small neighbourhood of pixels/area because big neighborhood pixels goes away and on the other hand, small areas are more robust to occlusion and clutter.
A good feature should be one that changes its gradients in the most directions – so, that means that any translation from it in a search will change the pattern underneath it and we can know exactly where it is in the other image because when we get to the correct location there is a distinct pattern.
