DoG stands for Difference of Gaussian, which is proposed by D. Lowe. Our implementation is fully based on description
in . Comparing with D. Lowe's version, we optimize the parameter settings based on our experiment data set. It is the fastest detector in our toolkit since it is avoid of calculation of second order derivatives.
LoG stands for Laplacian of Gaussian, which is discussed in . Our implementation is motivated by the description in
paper [2, 4]. Basically, we choose the points which attain local maxima in its spatial
and scale spaces. Comparing with DoG, it runs roughly 2 times slower. However, based
on our observation, It has better localization accuracy than DoG.
We implement Harris-Laplacian based on description in . For one detected local interest point, it attains local
maxima with respect to its Harris function, meanwhile it also reaches local
maxima at its a charateristic scale according to LoG function.
- Hessian-Laplace Hessian-Laplacian is quite similar to Harris-Laplacian. The only difference is that, we replace Harris matrix with Hessian matrix for local maxima detection in the spatial domain.
SIFT is proposed by D. Lowe in . It has been proved to be useful for various
tasks such as object classification, panorama generation, wide baseline matching
and ND image identification. Our implementation is fully followed the description
F-SIFT is a flip invariant SIFT feature proposed by us. Basically, it achieves flip invariance while maitaining
similar performances as SIFT when other transformations such as scale, rotation and lighting changes in presence. For details,
users are referred to our recent technical report on this novel feature descriptor. The flip invariance of this feature is
demonstrated in the next section.
PCA-SIFT is proposed by Yanke . It's an another variant of SIFT. Instead of quantizing
the gradient field, PCA-SIFT maps the gradient field to a vector (typically 36 dimension).
Before conducting the mapping, the local patch is normalized to fixed size (e.g.
41x41). For LIP-VIREO, we fully incorporate the implementation provided by
Yanke. For integrity, we keep all its default settings.
Motivated by PCA-SIFT, we apply PCA mapping directly on SIFT feature. In addition,
each dimension of the feature vector is further normalized by its variance respectively.
This results in a more concise representation of the local patch. Specifically,
we name it as PSIFT. The PCA mapping matrix is trained based on more than 50,000
local patches. Currently we only provide feature vector with 36 dimension. PSIFT
has been tested for near-duplicate detection , and performs very well.
- Steerable filters
Steerable Filters are also known as local jet or textons. They are actually partial
derivatives obtained based on series of Gaussian kernels. We follow the implementation
described in  in which a 41X41 Gaussian window circles around the keypoint is
employed. We calculate the derivatives up-to the forth order. The derivatives plus
one pixel value 'v' are ordered as following.
v dx dy dxx dxy dyy dxxx dxxy dxyy dyyy dxxxx dxxxy dxxyy dxyyy dyyyy
- Matched by HarLap+SIFT
- Matched by HarLap+F-SIFT
 D.Lowe, Distinctive Image Features from Scale-Invariant Key Points. IJCV, vol. 60, pp. 91-110, 2004.
 Linderberg, Feature detection with automatic scale selection. IJCV, vol.~30, no.2, pp. 79 -116, 1998.
 Y.Ke and R. Sukthankar, PCA-SIFT: A More Distinctive Representation for Local Image Descriptors. CVPR, vol. 2, pp. 506-513, 2004.
 K.Mikolajczyk and C. Schmidm Scale and Affine Invariant Interest Point Detectors. IJCV, 60 (2004), pp. 63-86.
 C. Schmid, R. Mohr, C. Bauckhage, Evaluation of Interest Point Detectors. IJCV, 2000.
 Wanlei Zhao A Comprehensive Study over Flip Invariant SIFT. Technical report [pdf], May, 2011.
 K.Mikolajczyk and C. Schmid. A Performance Evaluation of Local Descriptors. IEEE TPAMI, vol. 27, no.10, pp. 1615-1630, 2005.