sift_pyocl is an implementation of SIFT algorithm on GPU, in OpenCL programming language.

The documentation can be found on the page of the SILX project.

Features :

• Can run on GPU and CPU
• SIFT descriptors are accessible as numpy array
• Helper for images features with affine transformation
• Compatible with the ipol implementation

This module is now part of the SILX project (silx.image.sift). All the maintenance efforts are focused in this repository. The current repository will nevertheless be occasionally updated, in order to provide an independent SIFT module.

Once downloaded, the module can be installed with

python setup.py

or locally with python setup.py install --user.

sift_pyocl provides various plans for the image alignment process : descriptors computation, descriptors matching and image alignment. For more informations, please refer to the documentation.

The SiftPlan object provides a plan for computing the SIFT descriptors of a given image. The plan is defined for a given image shape (and target device). Returning the SIFT descriptors from an image gives an extended flexibility for further processing (outliers removal, alignment, classification).

import sift_pyocl as sift
siftPlan = sift.SiftPlan(img.shape, img.dtype, devicetype="GPU")
kp = siftp.keypoints(img)

The MatchPlan objects provides a plan for matching the descriptors of two images.

matchPlan = sift.MatchPlan(devicetype="GPU")
kp = siftp.match(kp1,kp2)

The LinearAlign class provides a plan for aligning images deformed with linear transformations (translation, rotation, scaling, shear).

alignPlan = sift.LinearAlign(img1, devicetype="GPU")
img2_aligned = alignPlan.align(img2)