The feature points detection is needed in a lot of tasks of image processing, for example in image stitching, in object recognition or in content based retrieval. The method called SIFT (Scale-Invariant Feature Transform) has received wide reputation. (See "Distinctive image features from scale-invariant keypoints" by Dawid Lowe http://www.cs.ubc.ca/spider/lowe/home.html for details). AIBO's vision system uses the SIFT algorithm to recognise its charging station (http://en.wikipedia.org/wiki/Aibo). Also this algorithm is used to stitch images (http://autopano.kolor.com for example).
During the project of automatic scanned images stitching with my participation I proposed the enhanced variant of feature points detector more stable to noise and blur.
Test image Suggested SIFT
Test image 1 Test image 2
Suggested SIFT
To compare these two methods I realized the interface to show feature points on images. You can download the code for testing. Attention! Because of using the Matlab, the application runs for a long time so I resample the images to 512x512. While using color images the application converts them to grayscale first.
I'll appreciate any feedback and examples of bad results of suggested algorithm.
Source code for my binaries can be obtained for research purposes. Please contact me for details.