SIFT_PyOCL uses the following programming languages:
The project is hosted by GitHub: https://github.com/kif/sift_pyocl
Which provides the issue tracker in addition to Git hosting. Collaboration is done via Pull-Requests in github’s web interface.
In addition to the run dependencies, SIFT_PyOCL needs an OpenCL compiler.
python setup.py build test
Test suites ensure both non regression over time and ease the distribution under different platforms. As OpenCL compilers are different for each platform (AMD, Intel, Nvidia), it is crucial to check if the process is done correctly.
SIFT_PyOCL comes with test-suites for different stages of the process. All the OpenCL kernels (23 in total) are tested and compared to a Python implementation. For the last steps (keypoints orientation assignment and descriptor computation), the Python implementation is slow ; the ASIFT version can be used for the comparisons (see https://github.com/kif/imageAlignment/tree/numpy).
To run all the tests, use python test/test_all.py.
| Name | Stmts | Miss | Cover |
|---|---|---|---|
| sift/__init__ | 7 | 7 | 100% |
| sift/alignment | 29 | 0 | 0% |
| sift/match | 132 | 31 | 23% |
| sift/opencl | 136 | 93 | 68% |
| sift/param | 6 | 5 | 83% |
| sift/plan | 412 | 42 | 10% |
| sift/sift | 1 | 0 | 0% |
| sift/utils | 47 | 17 | 36% |
| test_algebra | 120 | 106 | 88% |
| test_all | 35 | 34 | 97% |
| test_convol | 156 | 107 | 68% |
| test_cuda_shm | 43 | 0 | 0% |
| test_gaussian | 119 | 95 | 79% |
| test_image | 151 | 121 | 80% |
| test_image_functions | 289 | 223 | 77% |
| test_image_setup | 74 | 67 | 90% |
| test_keypoints | 179 | 0 | 0% |
| test_keypoints_old | 147 | 71 | 48% |
| test_matching | 89 | 78 | 87% |
| test_preproc | 276 | 195 | 70% |
| test_reductions | 88 | 79 | 89% |
| test_transform 141 | 0 | 0% | |
| utilstest | 151 | 70 | 46% |
| TOTAL | 2677 | 1371 | 51% |
Each test file in the test/ folder aims at testing an OpenCL kernels files. For example, test/test_image.py is a test suite for openCL/image.cl. All the OpenCL kernels have a reference implementation in Python, in file test/test_image_functions.py – that is, SIFT is almost entirely re-written in Python, however, the functions are not written to be efficient.
For the last steps (orientation assignment, descriptors computation, matching), the Python implementation becomes slow because all the previous functions have to be called. Instead, one can use the feature module available here.
It contains a Python wrapper for a SIFT C++ implementation. After installing it (python setup.py build_ext --inplace), you can get the keypoints with
import feature, scipy.misc
l=scipy.misc.lena().astype("float32")
s = feature.SiftAlignment()
u = s.sift(l);
This module can be used to compare SIFT_PyOCL with feature.SiftAlignment results. SIFT_PyOCL was based on this implementation.
In the file test/test_image_setup.py, SIFT parameters can be modified. The tests are run on a single scale of one octave, this can be modified as well.
| Parameter name | Default value | Description |
|---|---|---|
| border_dist | 5 | Distance to the border. The pixels located at border_dist |
| peakthresh | 255.0*0.04/3.0 | Threshold for the gray scale. Pixels whose grayscale is below will be ignored. |
| EdgeThresh | 0.06 | Threshold for the ratio of principal curvatures when testing if point lies on an edge |
| EdgeThresh0 | 0.08 | Threshold for the ratio of principal curvatures(first octave) |
| doubleimsize | 0 | The pre-blur factor is \(\sqrt{\sigma_0^2 - c^2}\) with c = 0.5 if doubleimsize = 0, 1.0 otherwise |
| initsigma | 1.6 | Initial blur factor (standard deviation of gaussian kernel) |
| nb_keypoints | 1000 | Maximum number of keypoints, for buffers allocating |
| ocsize | 1 | Initially 1, then twiced at each octave. It is a power of two |
| scale | 1 | scale can be 1, 2 or 3. Any other value is invalid ! |
Additionally, the test image can be modified. Default is l2 = scipy.misc.lena().astype(numpy.float32). You can also specify the device to run on, at the bottom of test/utilstest.py : ctx = ocl.create_context("GPU"). Simply remplace “GPU” by “CPU” will run all the tests on the CPU.
The test suites files can have the following constant defined at the top of the file.
| Constant name | Description |
|---|---|
| SHOW_FIGURES | If True, displays the figures with matplotlib |
| PRINT_KEYPOINTS | If True, displays parts of the keypoints vector for debugging |
| USE_CPU | If True, runs the tests on CPU |
| USE_CPP_SIFT | If True, uses feature module for keypoints comparison instead of python |
To fasten the tests, you can choose octsize = 4 and scale = 1 for example, as there are certainly less keypoints found in the superior octaves.