SIFT: reusing some claims and replacing others by strictly different nonpatented methods?

Question

SIFT is a patented method that does both feature detection and description (US6711293 - http://www.google.com/patents/us6711293). The method consists of roughly of three points (please let me know if I am inaccurate):

1. Detected features are local extrema of Difference of Gaussians in scale-space.
2. Each detected feature are assigned (possibly) multiple dominant gradient orientations.
3. For each detected feature and for all possible dominant orientations, a descriptor is computed in a rotationally invariant manner from several histograms of gradients.

Can I be sued...

if I design another method which:

• reimplements exactly claim (1) and claim (2), i.e., including all the associated technical details,
• but, for which, claim (3) is replaced by a strictly different nonpatented method, e.g., DAISY, FREAK...
• and will be used in a commercial application? (sold anywhere in the world)

Do I still have to replace claims (1) and (2) by strictly different methods?

Answer 1

First lets get terminology straight, each of the long sentences that is preceded by a number is a claim. This patent has 20 of them. Each sub-aspect of a claim is called an element or a limitation of that claim.

Each numbered claim is like its own patent. Patents as a whole are not infringed, claims are infringed. To be in the clear, whatever you are doing needs to not be describable by any of the 20 claims. Now, to infringe a particular claim, you need to do everything that claim requires. To infringe a method comprising the steps a, b and c you would need to do a, b and c. (Somewhat simplified).

In this case there are three independent claims: 1, 10 and 19. If you avoid those you avoid all of the others, by defininition, since the others are all narrowed versions of one of the independent claims. Claim 1 requires the steps of locating, comparing, comparing, comparing and finally producing descriptors for a plurality of subregions. To not practice claim 1 you need to accomplish a result that serves the purpose of the descriptors without doing at least one of those steps.

Gaussian and histogram are not specified anywhere in the claim language. To even get to "blur" you need to go to claim 3. That means avoiding Gaussian operations, histograms or even blurring will not get you past claim 1, for example.