Template talk:ScaleSpaceNavbox

From Wikipedia, the free encyclopedia

I saw that the SIFT descriptor was inserted as one of the items under the header Feature detection. However, I'm not fully positive about this. In a categorization of these subjects, I think that the SIFT descriptor is much more suitable under a heading of feature descriptor, and then on an equal footing as the SURF and GLOH, as done in the Feature detection navigation box. There are also other feature descriptors, such as N-jets, global or regional receptive field histograms, which are not yet described in Wikipedia, but would be appropriate to develop in an article about feature description in relation to scale-space.

If one would mention the SIFT descriptor in this navigation box for scale-space articles, it would be necessary to also mention other feature detectors on an equal basis, such as the Laplacian of Gaussian and Difference of Gaussians which it is very closely related to, as well as the determinant of the Hessian, not to mention the various corner detectors that are based on scale-space operations. Tpl (talk) 09:16, 26 February 2008 (UTC)

Tony, I'm not so familiar with any of these, but my impression was that SIFT in particular is probably one of the most widely used and successful applications of the scale-space concept, partly because it includes the idea of optimization of feature strength across the scale dimension, not so much for the particular representation it uses at the found points of interest. Do the other methods do that, too? Dicklyon (talk) 16:02, 26 February 2008 (UTC)

Dick, I agree that the SIFT descriptor is today among the most successful and mostly used applications of scale-space theory to computer vision problems. The initial feature detection step in the SIFT algorithm, referred to as keypoint detection in Lowe's paper, however is based on difference of Gaussians and seen as a pure feature detector (in the orthodox interpretation of this concept as described in the article on feature detection (computer vision) it is not particularly unique; rather it is very closely related to other scale invariant feature detectors with automatic scale selection, such as the Laplacian of Gaussian and the determinant of the Hessian. The really main contribution of the SIFT descriptor as I see it is in the feature description stage, where David Lowe proposed to compute local position-dependent histograms and how he furthered this notion to an efficient and operational scheme for recognizing real-world objects in real-time, by partially combining a number of existing tools and then also providing novel research contributions where needed in term of efficient matching and recognition as well as integration with other higher-level vision tasks. From this view-point, I find it more appropriate to categorize the SIFT descriptor as a multi-scale feature descriptor.

Concerning the navigation box, the current scale-space navigation box points to the feature detection pages. From each one of these articles, there is a feature detection navigation box, which contains a link to the SIFT article, from what I find as an appropriate categorization, where the SIFT descriptor is also on an equal footing as the SURF and GLOH descriptors, and also the other main feature detectors within the scale-space framework are listed. An alternative approach, which I'm less faviour of, would be to develop the scale-space navigation box to be more like the current feature detection navigation box. Tpl (talk) 16:43, 26 February 2008 (UTC)