Summary

A self-contained python wrapper for vlfeat. It is based on libvl 0.9.9 (and should be adapted to run along with the current vlfeat by some kind person :-) ).

Requirements

• numpy
• Boost.Python
• matplotlib

OSX

Homebrew gives you boost via brew install boost.

setup.py should work for both a homebrew based and a macports based system. I need to run

ARCHFLAGS="-arch x86_64" python setup.py install

for a successful build.

Linux

Depending on your system, you may need to change

LinkArgs.append('-lboost_python-mt-py26')

in setup.py. Install boost.python on Ubuntu systems via aptitude install libboost-python-dev.

SIFT

Only the vl_sift (and __init__.py) module was changed a bit. Assuming you have box.pgm in your working directory, then

import numpy, Image, vlfeat
box = Image.open("box.pgm")
box_array = numpy.array(box)
[f, d] = vlfeat.vl_sift(box_array.T, first_octave=-1, verbose=True)

produces nearly identical descriptors (in d) to bin/sift from vlfeat. Note that

• you don't need to specify dtype=np.float32 when providing the box_array to vl_sift,
• you need to use box_array.T
• the first octave is by default 0 (in bin/sift this is set to -1).
• the x and y positions in f are swapped (compared to the binary),
  the first number refers to row in box_array and the second one to column.

The produced descriptors match those from the binary nearly perfectly (in particular, the sequence of numbers is the same), sometimes some values are off by 1. I don't know why that is (vl_sift_calc_keypoint_descriptor produces two different descriptors on seemingly identical input). If some has an idea (or, more probable, I'm doing something stupid here), please let me know.