Merge pull request #8869 from hrnr:akaze_part1
authorJiri Horner <laeqten@gmail.com>
Wed, 21 Jun 2017 11:33:09 +0000 (13:33 +0200)
committerAlexander Alekhin <alexander.a.alekhin@gmail.com>
Wed, 21 Jun 2017 11:33:09 +0000 (14:33 +0300)
commit5f20e802d2a6802a26d18f7780db2bc0332cc394
tree2fb80e1369070aedeca713172e69e8d7f87e268f
parent437ca0b62a5daf1ebf39256b2a062500a66aab3b
Merge pull request #8869 from hrnr:akaze_part1

[GSOC] Speeding-up AKAZE, part #1 (#8869)

* ts: expand arguments before stringifications in CV_ENUM and CV_FLAGS

added protective macros to always force macro expansion of arguments. This allows using CV_ENUM and CV_FLAGS with macro arguments.

* feature2d: unify perf test

use the same test for all detectors/descriptors we have.

* added AKAZE tests

* features2d: extend perf tests

* add BRISK, KAZE, MSER
* run all extract tests on AKAZE keypoints, so that the test si more comparable for the speed of extraction

* feature2d: rework opencl perf tests

use the same configuration as cpu tests

* feature2d: fix descriptors allocation for AKAZE and KAZE

fix crash when descriptors are UMat

* feature2d: name enum to fix build with older gcc

* Revert "ts: expand arguments before stringifications in CV_ENUM and CV_FLAGS"

This reverts commit 19538cac1e45b0cec98190cf06a5ecb07d9b596e.

This wasn't a great idea after all. There is a lot of flags implemented as #define, that we don't want to expand.

* feature2d: fix expansion problems with CV_ENUM in perf

* expand arguments before passing them to CV_ENUM. This does not need modifications of CV_ENUM.
* added include guards to `perf_feature2d.hpp`

* feature2d: fix crash in AKAZE when using KAZE descriptors

* out-of-bound access in Get_MSURF_Descriptor_64
* this happened reliably when running on provided keypoints (not computed by the same instance)

* feature2d: added regression tests for AKAZE

* test with both MLDB and KAZE keypoints

* feature2d: do not compute keypoints orientation twice

* always compute keypoints orientation, when computing keypoints
* do not recompute keypoint orientation when computing descriptors

this allows to test detection and extraction separately

* features2d: fix crash in AKAZE

* out-of-bound reads near the image edge
* same as the bug in KAZE descriptors

* feature2d: refactor invariance testing

* split detectors and descriptors tests
* rewrite to google test to simplify debugging
* add tests for AKAZE and one test for ORB

* stitching: add tests with AKAZE feature finder

* added basic stitching cpu and ocl tests
* fix bug in AKAZE wrapper for stitching pipeline causing lots of
! OPENCV warning: getUMat()/getMat() call chain possible problem.
!                 Base object is dead, while nested/derived object is still alive or processed.
!                 Please check lifetime of UMat/Mat objects!
21 files changed:
modules/features2d/perf/opencl/perf_fast.cpp [deleted file]
modules/features2d/perf/opencl/perf_feature2d.cpp [new file with mode: 0644]
modules/features2d/perf/opencl/perf_orb.cpp [deleted file]
modules/features2d/perf/perf_agast.cpp [deleted file]
modules/features2d/perf/perf_fast.cpp [deleted file]
modules/features2d/perf/perf_feature2d.cpp [new file with mode: 0644]
modules/features2d/perf/perf_feature2d.hpp [new file with mode: 0644]
modules/features2d/perf/perf_orb.cpp [deleted file]
modules/features2d/src/akaze.cpp
modules/features2d/src/kaze.cpp
modules/features2d/src/kaze/AKAZEFeatures.cpp
modules/features2d/test/test_descriptors_invariance.cpp [new file with mode: 0644]
modules/features2d/test/test_descriptors_regression.cpp
modules/features2d/test/test_detectors_invariance.cpp [new file with mode: 0644]
modules/features2d/test/test_detectors_regression.cpp
modules/features2d/test/test_invariance_utils.hpp [new file with mode: 0644]
modules/features2d/test/test_rotation_and_scale_invariance.cpp [deleted file]
modules/stitching/perf/opencl/perf_stitch.cpp
modules/stitching/perf/perf_precomp.hpp
modules/stitching/perf/perf_stich.cpp
modules/stitching/src/matchers.cpp