文件列表:

Brief_Test\Brief_Test\BRIEF.h (23860, 2012-04-06)
Brief_Test\Brief_Test\BRIEFMatcher.h (2638, 2012-04-06)
Brief_Test\Brief_Test\Brief_Test.cpp (8514, 2012-04-06)
Brief_Test\Brief_Test\Brief_Test.vcproj (5554, 2012-04-06)
Brief_Test\Brief_Test\Brief_Test.vcproj.DICOM.Administrator.user (1407, 2012-04-06)
Brief_Test\Brief_Test\GroundTruth.h (4430, 2012-04-06)
Brief_Test\Brief_Test\MatchVerifier.h (2687, 2012-04-06)
Brief_Test\Brief_Test\MyKeypoint.h (1931, 2010-09-21)
Brief_Test\Brief_Test\stdafx.cpp (297, 2012-04-06)
Brief_Test\Brief_Test\stdafx.h (320, 2012-04-06)
Brief_Test\Brief_Test\targetver.h (765, 2012-04-06)
Brief_Test\Brief_Test\TestSampler.h (2096, 2012-04-06)
Brief_Test\Brief_Test\TestVisualizer.h (3529, 2010-09-21)
Brief_Test\Brief_Test\Thumbs.db (8704, 2012-04-06)
Brief_Test\Brief_Test\utils.cpp (7856, 2012-04-06)
Brief_Test\Brief_Test\utils.h (9907, 2012-04-06)
Brief_Test\Brief_Test\wall\H1to1p (18, 2010-09-21)
Brief_Test\Brief_Test\wall\H1to2p (178, 2010-09-21)
Brief_Test\Brief_Test\wall\H1to3p (177, 2010-09-21)
Brief_Test\Brief_Test\wall\H1to4p (178, 2010-09-21)
Brief_Test\Brief_Test\wall\H1to5p (179, 2010-09-21)
Brief_Test\Brief_Test\wall\H1to6p (177, 2010-09-21)
Brief_Test\Brief_Test\wall\img1.ppm (2100134, 2010-09-21)
Brief_Test\Brief_Test\wall\img2.ppm (1795215, 2010-09-21)
Brief_Test\Brief_Test\wall\img3.ppm (1795215, 2010-09-21)
Brief_Test\Brief_Test\wall\img4.ppm (1795215, 2010-09-21)
Brief_Test\Brief_Test\wall\img5.ppm (1795215, 2010-09-21)
Brief_Test\Brief_Test\wall\img6.ppm (1795215, 2010-09-21)
Brief_Test\Brief_Test.ncb (10660864, 2012-04-06)
Brief_Test\Brief_Test.sln (896, 2012-04-06)
Brief_Test\Brief_Test.suo (34304, 2012-04-06)
Brief_Test\Release\Brief_Test.exe (49152, 2012-04-06)
Brief_Test\Release\说明.txt (83, 2012-04-06)
Brief_Test\Brief_Test\wall (0, 2012-04-06)
Brief_Test\Brief_Test (0, 2012-04-06)
Brief_Test\Release (0, 2012-04-06)
Brief_Test (0, 2012-04-06)

BRIEF Demo Code Version 1.0 =============== michael.calonder@epfl.ch http://cvlab.epfl.ch/software/brief/index.php Thank you for your interest in BRIEF. This is an implementation that closely follows the version in [1]. The main difference between this implementation and that of [1] is that the former relies on integral images for computing the mean filter response of an image patch where the latter employs Gaussian filtering. Since the accuracy of the two has been proven identical, the mean filter is preferable over the Gaussian filter for its lower computational complexity. The code is released under the GNU General Public License V2. 1. BUILDING THE CODE -------------------- The only dependency is the OpenCV library which is freely available at [2]. Once installed, double-check that pkg-config can locate it. To test this, type $ pkg-config opencv --libs --cflags which should result in somehting similar to -I/usr/local/include/opencv -L/usr/local/lib -lcxcore -lcv -lhighgui -lcvaux. To build the code: 1. cd brief 2. make --> creates ./lib/libbrief.so 3. cd ../test_app 4. make --> creates ./main Note: For matching binary vectors, the POPCNT instruction from the SSE4.2 instruction set is particularly efficient and results in a tremendous speed-up. Intel CPUs starting from the i7 series support this instruction. Under Linux, POPCNT availability can be checked via $ less /proc/cpuinfo | grep sse4_2 If you find sse4_2 or alternatively the popcnt flag defined, you should edit the Makefile for the test application (test_app/Makefile) and uncomment -msse4.2. Be aware that doing so for a CPU without POPCNT will result in an Invalid Instruction. 2. RUNNING THE CODE ------------------- When executing the code, the system must be able to localize the library libbrief.so. For convenience, you can simply link that file into the test_app folder by typing [test_app]$ ln -s ../brief/lib/libbrief.so or, alternatively, you can modify the LD_LIBRARY_PATH environment variable to include the lib directory, $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:../brief/lib The demo code shows how to do three things with BRIEF: 1) Performing point matching and compute the rate of success. For example, $ ./main --match 3 will detect feature points in wall/img1.ppm and wall/img3.ppm, compute the corresponding BRIEF descriptors, match them and compute the percentage of correct matches. The output should look like [OK] Got 88.97% of 1306 retrieved matches right The implementation uses SURF points but you can use a point detector of your choice. 2) Measuring time required to build the descriptors. For example, typing $ ./main --time-desc 300 will evaluate the time required for computing around 2800 descriptors by averaging the result over 300 computations. The output I get is [OK] Computing 2835 descriptors took 13.430 ms (~ 2.426 ms/512 desc) 3) Measuring time required to match a set of binary vectors. For example, $ ./main --time-match 100 512 computes 512^2 = 262'144 distances and averages the result over 100 runs. The results in [OK] Matching exhaustively 512 vectors (3e+05 dist comp) took 12.846 ms on my machine which does NOT support the above POPCNT instruction. Clearly one could take advantage of fast bitcount implementations but doing so will hardly be faster than POPCNT--which will soon be widely available. In case of problems such as SEGFAULTs, remember to build and run the code in DEBUG mode which will enable a plethora of (expensive) run-time checks, helping you to locate the issue more quickly. 3. HARDCODED PARAMETERS ----------------------- There is a handful of hardcoded parameters inside the BRIEF class. DESC_LEN Descriptor length, or the number of tests. Default: 256 (i.e. BRIEF-32). INIT_PATCH_SZ The image patch size around a keypoint in which the tests are defined. The "INIT" prefix indicates that the patch size can vary as soon as the rotateTests() and scaleTests() functions are used. Default: 48. INIT_KERNEL_SZ The kernel size of the mean filter in pixels used to smooth the patch. The actual kernel size can differ when scaleTests() is called, causing the kernel to rescale. Default: 9. USE_INTEGRAL_IMAGE If 1, we pre-compute an integral image allowing for fast smoothing. The speed-up is notable for small and medium-sized images. Default: 1. SMOOTH_ENTIRE_PATCH (defined iff USE_INTEGRAL_IMAGE==1) If 1, the entire patch is smoothed before the tests are applied. However, if the number of tests is small compared to the patch area, it is usually more efficient to smooth only around the test locations rather than the entire patch. Default: 0. SOS_ORI If your feature point detector provides an orientation, BRIEF can rotate the tests into a canonical orientation before using them, allowing the descriptor to become rotationally invariant. BRIEF's orientation is 0 at three o'clock and varies between -PI and +PI. To use the functionality, you will have to extend interpretScale() to correctly map the orientation of your detector to [-PI,PI]. For illustration, we include this done for OpenCV's SURF implementation which can be enabled by setting SOS_ORI to SOS_SURF_OPENCV. Default: SOS_DO_NOT_INTERPRET. SOS_SCALE Same as SOS_ORI but for scale. Enabling the experimental features SOS_ORI and SOS_SCALE can cause one or more tests to fall outside the image region which potentially causes SEGFAULTS. This can be prevented by compiling the code in DEBUG mode since this enables checks that allow to skip tests outside the image area. 4. PLEASE NOTE -------------- A note regarding the fraction of correctly retrieved matches: The values are NOT directly comparable with those in [1], where a similar measure known as "recognition rate" was employed. The reasons are: * Here we match 1000-1500 points where in [1] the default number of matches was 512. * Here the method has a way to reject matches using a left-right consistency check. In [1], the method had to establish exactly one match for each feature point. In other words, [1] required the methods to match two same-sized sets of features, from which outliers have been removed. Here we allow outliers in both sets. Nevertheless, neither of the methods performs checks for geometric consistency. The code has only research quality and is therefore not particularly optimized for speed. Suggestions for improvements are most welcome. The test data in the test_app/wall folder is publicly available from http://www.robots.ox.ac.uk/~vgg/research/affine and has been included without modifications. 5. WINDOWS AND MAC OSX ---------------------- This version does not provide an automated way to build the code under Windows or Mac OSX, though it should be fairly easy to do so: The source code is standard C++ and depends on OpenCV only, which is cross-platform [2]. 6. REFERENCES ------------- When referring to BRIEF, please cite [1]. [1] @inproceedings{Calonder10-brief, author = "M. Calonder and V. Lepetit and C. Strecha and P. Fua", title = {{BRIEF: Binary Robust Independent Elementary Features}}, booktitle = "European Conference on Computer Vision", month = "September", year = 2010 } [2] http://opencv.willowgarage.com

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：