FREAK Image Features

Dan Bell
May 2, 2016

Image Key Points and Features

In Computer Vision and Image Processing, image features have become a powerful tool with a variety of uses. Object tracking, image recognition and 3D Reconstruction are just three of such examples of modern feature usage. An image key point is a point in an image that has enough intensity variation that it can be tracked reliably between images.

The ARToolkit relies on FREAK features for tracking and orientation estimation processes, which are a binary descriptor pattern that is fast and reliable.

Keypoint Examples

Binary Features versus Floating Point

A binary feature pattern has the following three requirements.

  • A sampling Pattern
  • Orientation Compensation
  • Sampling Pairs

The sampling pattern is a process of selection as to where to sample points in the region around the descriptor. The orientation compensation is some mechanism by which to measure the orientation of the key point and rotate it to compensate for rotation changes when matching. Finally, sampling pairs are the selected sequence of pairs to compare when building the final descriptor.

Floating point approaches can use more robust tricks to create feature points, including sampling gradients around a key point as with SIFT(Scale Invariant Feature Transform) or SURF(Speeded Up Robust Features).

The ARToolkit uses FREAK features, Fast REtinA Keypoints.

Sampling pattern: The sampling pattern is a series distributed circles that grow sparser towards the edge of the pattern, similar to the distribution of photoreceptors on the human retina.

Sampling Patterns

Orientation Compensation: Orientation is calculated using 45 symmetric pairs from the centre. It has larger steps than that of BRISK and thus needs lower memory.

The sampling pairs can be any of the selected sample points from the retina, provided they are used consistently along the pipeline.

Performance

Comparing feature features to other modern methods, it becomes clear that FREAK features out perform many of the conventional approaches.

1) The number of features detected. Feature Detection Graph

2) The performance of re-detection or matching for FREAK features. Performance Graph