Direct Geometric and Photometric Image Registration

Adrien Bartoli >> Research >> Direct Geometric and Photometric Image Registration

Image registration consists in estimating geometric and photometric transformations that align a template and an image as best as possible. The direct approach consists in minimizing the intensity discrepancy between the aligned template and image. The inverse compositional algorithm has been recently proposed by Baker et al. [1] for the direct estimation of groupwise geometric transformations. It is efficient in that it performs most computationally expensive calculations at the pre-computation phase.
    We propose the gain and bias inverse compositional algorithm which estimates, along with the geometric transformation, a photometric one modeling for example global lighting change. Our algorithm preserves the efficient pre-computation-based design of the original inverse compositional one. Baker et al. proposed the simultaneous inverse compositional algorithm in [2], which deals with linear appearance variations, but does not preserve the advantage of the inverse compositional algorithm. Our gain and bias inverse compositional algorithm performs exactly the same calculations, but requires less computational time. More recently, we proposed the dual inverse compositional algorithm which handles general, groupwise photometric transformations and color images.
    Experimental results on simulated and real data show that with our fairly optimized implementation in Matlab, the computational cost is reduced by factors of at least 2 for our algorithm.

Example template (left) and image (right) to be registered

Iteration 1 - Intensity error 55.41 Iteration 10 - Intensity error 34.56 Iteration 19 - Intensity error 28.22 Iteration 28 - Intensity error 4.13

Papers

Groupwise Geometric and Photometric Direct Image Registration
A. Bartoli
IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 30, No. 12, pp. 2098-2108, December 2008.

[PDF]

Groupwise Geometric and Photometric Direct Image Registration
A. Bartoli
BMVC'06 - Proceedings of the Seventeenth British Machine Vision Conference, Edinburgh, UK, September 2006.

[PDF]

Direct Image Registration With Gain and Bias
A. Bartoli
Invited paper at the Topics in Automatic 3D Modelling and Processing Workshop, Verona, Italy, March 2006.

[PDF]


Code -- new: see here for code download

The DIRT (Direct Image Registration Toolkit) Matlab library implements dual inverse compositional algorithm for homographic warps. Example images and functions are included. Gray-level and color images are handled. The method was originally proposed in my BMVC'06 paper and is now published in my PAMI'08 paper - please cite it if you use the DIRT package! The new 1.1 and 1.2 versions correct several bugs and have an experimental point matching procedure based on patch alignment.

v1.2 2009.01.09 DIRT_v1p2.rar - new motion models: 'Affine' (6 parameters) and 'Rt' (3 parameters)
- checking of the GN-Hessian's condition number
- 2009.08.14: bug fixed for debug2 mode used with color images
v1.1

2006.10.06

DIRT_v1p1.rar - new motion model: '3PtHomography' (6 parameters)
- experimental point matching procedure
v1.0

2006.06.29

DIRT_v1p0.rar  

The cDIRT library is a C++ version of DIRT contributed by Pierre Georgel and packaged by Ludovic Magerand. It uses OpenCV, and has been tested on Linux and Windows.

v1.0

2009.02.13

cDIRT_v1p0.rar - initial version, based on DIRT v1.1

The DirectH Matlab library implements the gain and bias inverse compositional algorithm for homographic warps. Example images and functions are included. Only gray-level images are handled. The method is described in my paper for the 2006 workshop in Verona.

v1.1 2006.03.22 DirectH_v1p1.rar
v1.0

2006.03.08

DirectH_v1p0.rar