Viirya

"Distinctive Image Features from Scale-Invariant Keypoints"

By David G. Lowe

--

This paper proposes the Scale Invariant Feature Transform (SIFT) approach for image matching problem. It is widely used in computer vision and image retrieval related fields.

It is a method to extract local interest points from images. Importantly, these feature points are invariant to image scale, rotation. It includes four main stages:

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

*Scale-space extrema detection

The method detects keypoints using a cascade filtering approach that identifies candidate locations that are then examined in further. To detect keypoints invariant to scale change, it can be accomplished by searching for stable features across all possible scales.

SIFT uses the difference-of-Gaussian function to obtain such locations of candidate keypoints.

*Keypoint localization

Each candidate keypoints from previous step is compared to its neighbors in same image and other neighbors in images of the scale above and below.

*Orientation assignment

In this step, SIFT assigns a consistent orientation to each keypoint. With such orientation, the keypoint descriptor generated in next step can achieve invariance to image rotation.

*Keypoint descriptor generation

After assigning the location, scale and orientation to each keypoints in an image, SIFT generates keypoint descriptors based on the gradients in the regions around the keypoint locations.

Observed from the experiments by Lowe, 128-dim feature vector performs best. 128-dim SIFT feature becomes commonly used feature.