The python cached type query */

Check type equivalence in a name list like <name1>|<name2>|... Return 0 if equal, -1 if nb < tb, 1 if nb > tb */

Check type equivalence in a name list like <name1>|<name2>|... Return 0 if not equal, 1 if equal */

Compare two type names skipping the space characters, therefore "char*" == "char *" and "Class<int>" == "Class<int >", etc. Return 0 when the two

Search for a swig_type_info structure for either a mangled name or a human readable name. It first searches the mangled names of the types, which is

Return the best assignment (= list of correspondences) for a given maxima from a pyramid top, this function returns a list of weigthed correspo

average-pool in 2x2 px non-overlapping cells */

Compute the dx,dy gradient on the image based on a [-1,0,1] mask. =0 : no prior smoothing =1 : sobel smoothing */

Compute the Histogram of oriented gradient for each pixel. Number of orientations is determined by hog->tz; method determines orientation bining

compute original, unsmoothed, gradient */

compute gradient smoothed with Sobel mask */

matrix-matrix multiplication with several SGEMM (each is single-threaded) res = dot(patches, convolved_hog) P*npix P * nh nh * npix */

convolution of each patch within a local neighborhood ngh_rad = max translation neighborhood has size 2*ngh_rad patch at (x,y

Return a list of local maxima in the scale-space of scores */

joint max-pooling and subsampling */

Prepare a grid of cell positions in the first image for a given scale. Big cells inherit the cell at the previous scale. size = size of cells at c

Prepare image for dotprod : dot(patches, res) where patches is n_patches x patch_dim set outside of the image to be equal to (0,...,ninth_val) *

Sample a set of patches from a HOG image. grid : array of (x,y) position of the patches size: size of the patches, ie. [x,x+size[ x [y,y+size[

Compute the (sparse) convolutions specified by <children> on <map> and put the result in <res>. A standard order is assumed on the children: a

Subsample an array, equivalent to res = trueimg[:,offset_y::2,offset_x::2] except at boundaries, where the rules are a bit more complex: if img

Compute the multi-scale pyramid response

extract pixel descriptor for both images

Compute the correlation of all patches with the second image (hog). In case of ngh_rad>0, the correlation is only computed in a small local neighbo

return points corresponding to patch matches

this function gather correspondences from each local maximum in the response maps */

crop an image (in-place) */

convert a float image to a consecutive array a bit stupid but well */

merge correspondences from several rotated/scaled version of an image into a single common reference frame rot0 = 2x3 rotation matrix: (pt in ro

Normalize weights in border areas of width <gap>. There are 9 areas: top-left, top-middle, top-right, ..., bottom-right. sum_divf indicates the

correct the convolution on the boundaries of the image ttx, tty: true shape of the res_map (in case of using offsets) */

reallocate the memory of an image to fit the new width height */

rotate a descriptor HOG image by a given angle

set default params