[paper reading]"Distinctive Image Features from Scale-Invariant Keypoints," Lowe, IJCV, 2004

This paper introduce a method called Scale Invariant Feature Transform (SIFT) to find local feature points on a images which are invariant to image scale and rotation. There are four steps in this method, which are :

1.Scale-space extrema detection
2.Keypoint localization
3.Orientation assignment
4.Keypoint descriptor

The first two steps are using to find locations of the interesting points on images, and than use the third and forth steps to find descriptor of the interesting points to make it invariant to affine transformation and also highly distinctive when use to matching. This method is a cascade filtering approach, which expensive operation such as descriptor computation only apply at location that pass an initial test, to avoid the expensive computation time. After introduce the detail of the SIFT method, this paper then shows the application to object recognition. Although the keypoint descriptor find using SIFT method is highly distinctive, in a cluttered image, a lot of feature from background will have false match since there are no correct match for those feature in database. Therefore, this paper use a subset of feature points that agree on location, scale and rotation in matching image to filter the desire feature points and reject the outlier and hence improve the result of the object recognition. For feature point matching, this paper use an approximate algorithm called Best-Bin-First algorithm (Beis and Lowe, 1997), to find approximate nearest neighbor in short time. Author of this paper provide a demo program to show how the algorithm works, and the result is really amazing. This is a classic paper in the field which is cited more then six thousand times according to google scholar!!