Bookmark and Share

Computer Vision Metrics: Chapter Four (Part C)

Register or sign in to access the Embedded Vision Academy's free technical training content.

The training materials provided by the Embedded Vision Academy are offered free of charge to everyone. All we ask in return is that you register, and tell us a little about yourself so that we can understand a bit about our audience. As detailed in our Privacy Policy, we will not share your registration information, nor contact you, except with your consent.

Registration is free and takes less than one minute. Click here to register, and get full access to the Embedded Vision Academy's unique technical training content.

If you've already registered, click here to sign in.

See a sample of this page's content below:


For Part B of Chapter Four, please click here.

Bibliography references are set off with brackets, i.e. "[XXX]". For the corresponding bibliography entries, please click here.


Descriptor Density

Depending on the image data, there will be a different number of good interest points and features, since some images have more pronounced texture. And depending on the detector method used, images with high texture structure, or wider pixel intensity range differences, will likely generate more interest points than images with low contrast and smooth texture.

A good rule of thumb is that between 0.1 and 1 percent of the pixels in an image can yield raw, unfiltered interest points. The more sensitive detectors such as FAST and the Harris detector family are at the upper end of this range (see Appendix A). Of course, detector parameters are tuned to reduce unwanted detection for each application.

Interest Point and Descriptor Culling

In fact, even though the interest point looks good, the corresponding descriptor computed at the interest point may not be worth using and will be discarded in some cases. Both interest points and descriptors are culled. So tuning the detector and descriptor together are critical trial-and-error processes. Using our base assumption of 0.1 to 1 percent of the pixels yielding valid raw interest points, we can estimate the possible detected interest points based on video resolution, as shown in Table 4-2.


Table 4-2. Possible Range of Detected Interest Points per Image

Depending on the approach, the detector may be run only at mono-scale or across a set of scaled images in an image pyramid scale space. For scale space search methods, the interest point detector is run at each pixel in each image in the pyramid. What methods can be used to cull interest points to reduce the interest point density to a manageable number?

One method to select the best interest points is to use an adaptive detector tuning method (as discussed in Chapter 6 under “Interest Point Tuning”). Other approaches include only choosing interest points at a given threshold distance apart—for example, an interest point that cannot be adjacent to another interest point within a five-pixel window, with the best candidate point selected within the threshold.

Another method is to vary the search strategy as discussed in this...