ECC Image Alignment Algorithm

This webpage provides the code of the ECC image alignment algorithm (for details see [1],[2]). This algorithm can be used in a wide range of applications such as image registration, mosaicing, object tracking, super-resolution, visual surveillance by moving cameras.

In short, ECC image alignment algorithm estimates the geometric transformation (warp) between the input and the template image and it returns the warped input image which must be close to the template one, as it is shown below. The estimated trasnformation is the one that maximizes the correlation coefficient between the template and the warped input image.

Find please below both Matlab and OpenCV implementations of the ECC algorithm. Note that the code is provided "as is" without any kind of warranty. The use and the redistribution of the code is only permitted for academic purposes.

News - Updates

>> Update 13-May, 2012:
ECC deals with most of 2D planar transformations, i.e. translation, euclidean, affine, homography. Both versions (Matlab, OpenCV) are compatible with any type of the above transformations.

>> Update 14-Mar, 2013:
Bugs in Matlab version for partially overlapped images and Euclidean case have been fixed.

>> 1-Mar, 2013:
The ECC algorithm is a new function in OpenCV library, findTransformECC(). Get the latest trunk of the library from here. The documentation of the current trunk of the lirbary is available here . The function will be also included in the next official release of the library (2.5)

>> Please cite the papers even if you are using ECC algorithm through the OpenCV library.

Matlab Code

A multiresolution execution based on gaussian pyramids is optional through this implementation. Except for the comparing images, the user sets as inputs the number of pyramid levels (of course it can be 1), the number of iterations per level, the type of geometric transformation (translation, euclidean, affine or homography), and, optionaly, the warp initialization matrix. The latter is useful for speedup and when either a strong transformation applies and the algorithm is likely to be trapped in local optimum or the template corresponds to a specific region of the input image.

Download the Matlab code from here.
The current Matlab implementation is also available in Matlab Central.

Example of using ECC for 40 iteration in a single level (no multi-resolution):

>>img=imread('image.pgm'); >>tmplt=imread('template.pgm'); >>result=ecc(img, tmplt, 1, 40, 'homography');

for help type:

>>help ecc

or see the M-file ecc_demo.m.

Note that the real transform between the above images is a homography. Homographies are usually under-estimated using an affine or any other 2d model.

OpenCV Code

I encourage you to install the latest trunk of the OpenCV library which includes the ECC algorithm. Note that the implementation which is provided with the OpenCV library is updated and deals better with partially overlapped images.

If you want to use your old OpenCV trunk, you can find here a C/C++ implemetation of the ECC algorithm which is compatible with OpenCV 2.1 or newer versions. You can download the UPDATED source code. Otherwise, you can keep only the windows executable file. The required DLL files are included in the zipped file.

Acknowledgements: Many thanks to Mark Asbach (Fraunhofer IAIS) for his contribution to code refactoring. The algorithm is really fast because of Mark's skills!

Find under my BitBucket account another implementation of the algorithm that runs even faster since Jacobian matrix is computed by block processing! Download the Windows executable file. This version deals with affine and homography transformations.

The executable files have been developed using VC++ 2008 (express edition) and have been tested under WinXP 32bit and Win7 64bit platforms. You need to recompile the files if you are working on other OS.

Example of use:

X:> ecc -t template.pgm -i image.pgm -o warp.ecc -m homography

or run the executable file without arguments to see help and more options.

Note that after the execution, the file warp.ecc and the image warped.png will be written in the disk. The first contains the estimated warp matrix saved in YAML format (see below) and the second is the warped input image.

The warp initialization or the output file should have the following YAML format:

%YAML:1.0
Warp: !!opencv-matrix
   rows: 3
   cols: 3
   dt: f
   data: [ 1., 0., 0.,
       0., 1., 0.,
       0., 0., 1. ]

The above structure defines a 3x3 identity matrix with float values.

Run EXE file from Matlab

A matlab function that calls the above .exe file produced by the OpenCV code is also provided. Download it from here.

Example of calling this function (40 iteration are considered and the warp is initialized by the matrix written in the file init.txt):

>>[warped_image, warp] = ecc_opencv('image.pgm', 'template.pgm', 'homography', 'init.txt', 40);

You can ignore the last argument (the default value of iterations is 50). To ignore the initialization as well, give as input the zero value, i.e.

>>[warped_image, warp] = ecc_opencv('image.pgm', 'template.pgm', 'homography', 0);

If no initialization file is given, the warp is initialized by the identity transformation (identity matrix).

For help type

>>help ecc_opencv

or see the code of the file.

References

[1] G. D. Evangelidis , E. Z. Psarakis, "Parametric Image Alignment using Enhanced Correlation Coefficient Maximization", IEEE Transaction on Pattern Analysis & Machine Intelligence , Vol. 30, No. 10, pp. 1858-1865, 2008. (pdf)

[2] G. D. Evangelidis E. Z. Psarakis, "Projevtive Image Alignment by using ECC Maximization", in Proc. Int. Conf. on Computer Vision Theory and Applications (VISSAP), January 2008, Madeira, Portugal. (pdf)

Contact

For any bugs, questions or help, please contact the author
Georgios Evangelidis,
e-mail: firstname.lastname at inria dot fr