Switch branches/tags
Nothing to show
Find file History
Pull request Compare This branch is 15 commits ahead, 26 commits behind kif:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


Demo code for Affine-SIFT (ASIFT) image matching
Jean-Michel Morel (Jean-Michel Morel <morel@cmla.ens-cachan.fr>)
Guoshen Yu (yu@cmap.polytechnique.fr)

Version 2.2, April. 10, 2010

This directory contains the C++ code for ASIFT, a fully affine invariant image 
matching algorithm. 

************************** Unix/Linux/Mac Users **************************
For Unix/Linux and Mac users, the code is ready to be compiled. The 
executable adapted to your computer system is generated after compiling.

***************************** Windows Users ******************************
For Windows users, the executable as well as the code is provided. 
(For the executable, you need to download separately the installation file from 
http://www.ipol.im/pub/algo/my_affine_sift/ .)

**************************** Matlab Interface ****************************
Although the ASIFT program is standalone and can be executed without Matlab,
a Matlab interface is provided (for Unix/Linux/Mac/Windows users). 

Source code compilation and software usage across platforms is detailed in this
manual. If you have any problem using the this program, please contact Guoshen Yu 
For more information about ASIFT, please see the web page at 
You can also try ASIFT using the online demo. The online demo allows testing
ASIFT with your own images without installing the program. 

If you use the ASIFT code or software, please cite the following paper: 
J.M. Morel and G.Yu, ASIFT: A New Framework for Fully Affine Invariant Image 
Comparison, SIAM Journal on Imaging Sciences, vol. 2, issue 2, pp. 438-469, 2009. 



The source code needs to be compiled before the software can be used. 
The compilation requires the make program, and is typically straightforward.

- Library. 
This code requires the libpng library. You can automatically download, 
compile and include this library to the compiled program by adding the 
LOCAL_LIBS=1 option to the make commands.

- Image format. 
Only the PNG format is supported. 
Source compilation and software usage
1. Download the ASIFT code package and extract it. Go to that directory. 

2. Compile the source code (on Unix/Linux/Mac OS). 
There are two ways to compile the code. 
(1) RECOMMENDED, with Open Multi-Processing multithread parallelization 
(http://openmp.org/). Roughly speaking, it accelerates the program using the 
multiple processors in the computer. Run
make OMP=1

(2) If the complier does not support OpenMp, run 

If libpng (the official PNG reference library) is not installed in your computer, 
an option LOCAL_LIBS=1 should be added after make. Example
The compilation will automatically download and compile libpng and zlib and 
include the library to the program.

3. Run ASIFT. 
./demo_ASIFT imgIn1.png, imgIn2.png imgOutVert.png imgOutHori.png matchings.txt 

keys1.txt keys2.txt

-- imgIn1.png, imgIn2.png: Input images (in png format).
-- imgOutVert.png, imgOutHori.png: Output images (vertical/horizontal concatenated). 
The detected matches are connected by write lines.
-- matchings.txt: The file format starts with 1 integer giving the total number 
of matches. Then each line specifies the coordinates (col1, row1, col2, row2) 
of a pair of matched points. (col: horizontal axis, from left to right. 

row: vertical axis, from top to bottom.)
-- keys1.txt keys2.txt: ASIFT keypoints in the two images, in the same format 
as the SIFT keypoints of David Lowe. The file starts with 2 integers giving 
the total number of keypoints and the length of the descriptor vector for each 
keypoint (128). Then the location of each keypoint in the image is specified 
by 4 floating point numbers giving subpixel column and row location, scale, 
and orientation (in radians from -PI to PI). Finally, the invariant descriptor
 vector for the keypoint is given as a list of 128 integers in range [0,255]. 
-- [optional 0/1]. 1: input images resize to an area equal to 800x600 for ASIFT,
 in keeping the aspect ratio (by default). 0: no resize. The resize is to limit
 the ASIFT computation time. The results (output images, keypoint coordinates
 and scales) are normalized to the original image size, so the resize is 
"transparent" to the user. 

Example, run
./demo_ASIFT adam1.png adam2.png imgOutVert.png imgOutHori.png matchings.txt 

keys1.txt keys2.txt 

You get on the screen 
"WARNING: The input images are resized to 800x600 for ASIFT. 
         But the results will be normalized to the original image size.

Computing keypoints on the two images...
12928 ASIFT keypoints are detected. 
8972 ASIFT keypoints are detected. 
Keypoints computation accomplished in 24 seconds.
Matching the keypoints...
The two images match! 914 matchings are identified. log(nfa)=-1496.88.
Keypoints matching accomplished in 4 seconds."


For Windows users who do not want to recompile the source code, an ASIFT executable 
installation file can be download separately from 

- The provided Windows executable demo_ASIFT.exe has been compiled by the Intel C++ 
compiler on 32-bit Windows. It is executable on both 32-bit and 64-bit Windows, 
although it is not optimized for the latter. 

- The executable has not been extensively tested. If you have any problem using it,
please contact Guoshen Yu yu@cmap.polytechnique.fr. 

1. Download the installation file demo_ASIFTsetp.exe from 

2. Install the program.
Double click the file demo_ASIFTsetp.exe. A small library distributed by Microsoft 
(Microsoft Visual C++ 2010 Redistributable Package) will be installed to your PC. 
The ASIFT software will be installed to C:\Program Files\demo_ASIFT

3. Run ASIFT. 
Run a Dos command prompt (you find it in Start > All Programs > Accessories 

> Command Prompt)
- Go to the ASIFT directory by typing
cd C:\Program Files\demo_ASIFT.
- Run the ASIFT program by typing
demo_ASIFT adam1.png adam2.png imgOutVert.png imgOutHori.png matchings.txt 
keys1.txt keys2.txt
(It follows the same syntax of that for Unix/Linux/Mac described above.)

You can of course move the ASIFT directory C:\Program Files\demo_ASIFT to 
wherever that is more convenient. 

1. If you are able to run the program but the results are not written to 
the output files, check and make sure that you have the write file permission 
in the demo_ASIFT directory.

2. Microsoft Visual Studio is NOT required to run the program. However, 
in case you cannot run the program (for example some library or dll is missing), 
you may try installing a Microsoft Visual Studio and then running again 
the program. 

Windows users who want to recompile the source code themselves, please follow the 
guidelines below for a smooth compilation. 

It is assumed that the CMake and Visual Studio are installed. Moreover, 
the Intel C++ compiler, which supports vectorization, is highly recommended. 
For an optimized software execution speed, it is crucial that both OpenMP
and vectorization are activated and effective (see below). (Note that
the CPU must have SSE2 or higher instruction sets in order to make vectorization
possible. The /arch:SSE2 opiton must be turned on, which should have
been normally the case.)

1. If you don't have CMake, an open-source build system,
then download and install it from http://cmake.org

2. Launch the CMake program cmake-gui.exe

3. Provide the paths
- "Where is the source code :" (the path that leads to CMakeLists.txt and 
the ASIFT source code), e.g., E:/demo_ASIFT_src

- Create a subdirectory Build under demo_ASIFT_src

- "Where to build the binaries:" (the directory where build project and 
object files will be stored), e.g., E:/demo_ASIFT_src/Build

4. Press "Configure" and select "Use default native compilers" and 
select the code IDE that you have in your machine, e.g., Visual Studio 10.

5. Press "Generate". A Visual Studio solution file ASIFT.sln will
be created under E:/demo_ASIFT_src/Build

6. Launch Visual Studio 10, and open ASIFT.sln. 

7. Switch the mode from Debug to Release.

8. Not mandatory, but HIGHLY RECOMMENDED (for program acceleration).
- Switch from the Microsoft Visual C++ 
compiler to the Intel C++ compiler, if you have it. 

- Turn on OpenMP
Click the demo_ASIFT project in the Solution Explorer. Select the menu 
Project -> Property -> Configuration Properties -> C/C++ -> Language [Intel C++]
Set OpenMP Support to "Generate Parallel Code (/Qopenmp)"

- Open demo_lib_sift.cpp under demo_ASIFT/Source Files in the Solution Explorer.
Make the following code change in demo_lib_sift.cpp.
unsigned short distsq = 0; ----> int distsq = 0;

Note: This manipulation allows the compiler to vectorize the SIFT code
comparison, which accelerates the ASIFT keypoint matching by a factor of
 from 5 to 20. To have the vectorization possible, this distsq variable must
 be in unsigned short when compiled with make on Linux/Mac, and be instead
int when complied with the Intel C++ on Windows. Therefore 
this format should be adapted according to the compiler. (So don't
 forget to change it back if you want to compile the code on Linux/Mac.) 
This vectorization is not achieved with the Visual C++ compiler. 

9. Build solution. An executable will be created in Build/Release. 
Run it in a Dos command prompt. 
demo_ASIFT adam1.png adam2.png imgOutVert.png imgOutHori.png matchings.txt
 keys1.txt keys2.txt
(It follows the same syntax of that for Unix/Linux/Mac described above.)

Run ASIFT via Matlab: Open test_demo_ASIFT.m in Matlab and execute the script. 
The Matlab interface reads most standard image formats.

- The epipolar geometry filtering algorithm ORSA of Moisan and Stival is 
used at the end of the ASIFT algorithm 
to eliminate false matches. 

Pierre Moulon contributed to the SVD subroutine of ORSA. The SVD subroutine
the toolbox Eigen, which is LGPL-licensed. 

- We would like to thank Nicolas Limare for his help on developing the program, 
and Pierre Moulon for his help on making the Windows compilation much easier with 

Licensing conditions

This software is being made available for research purposes only. See the
 file LICENSE in this directory for conditions of use.