Research on Alignment/Tracking

Efficient compositional image aligment
Rationalizing efficient compositional image alignment. We study the issue of computational efficiency for Gauss-Newton (GN) non-linear least-squares optimization in the context of image alignment. We introduce the Constant Jacobian Gauss-Newton optimization, a GN scheme with constant Jacobian and Hessian matrices. We prove that the Inverse Compositional image alignment algorithm is an instance of this scheme. We also prove that the forward and inverse compositional algorithms are not equivalent.
Related Publications: IJCV'2015,
Fast homography estimation
Speeding-up homography estimation in mobile devices. We introduce a procedure for reducing the number of samples required for fitting a homography to a set of noisy correspondences using a random sampling method. This is achieved by means of a geometric constraint that detects invalid minimal sets.
Related Publications: JRTIP'2015,
3D nonrigid tracking
Efficient 3d nonrigid tracking. Efficient incremental image alignment is a topic of renewed interest in the computer vision community because of its applications in model fitting and model-based object tracking. We are working in efficient solutions to the 3D tracking of a head performing face expressions under changing illumination conditions.
Related Publications: VIE'2005, ICCV'2005 ICCV'2009
Illumination independent appearance-based tracking
Efficient appearance-based tracking with illumination changes and face expressions. separates facial expressions from illumination variations. The appearance of a face is represented by the addition of two independent linear subspaces modelling facial expressions and illumination. This simple model enables us to train the system with no manual intervention. We also introduce an efficient procedure for fitting this model, which can be used for tracking a human face in real-time.
Related Publications: BMVC'2006, ICPR'2006, IVC'2009
Some videos: [Video 1], [Video 2], [Video 3], [Video 4]

You can download the original sequences used in our tests (BMVC 2006 paper image sequences).

Efficient Eigentracking
Efficient appearance-based tracking We have developed an efficient way of minimizing the eigentracking for nonrigid motion estimation. It is based on the precomputation of motion templates to save on-line computation. It allows as to estimate appearance (PCA coefficients) and motion in real-time.
Related Publications: ANM'2004
Some videos: [Video 1]
SSD Based Tracking
SSD based 3D tracking. We have developed an algorithm for tracking a rigid object based on a piecewise planar model. The tracking is performed using a single incremental SSD-based tracker. The main feature of the approach presented is that it can track a rigid set of arbitrarily small patches all of which could not be individually tracked.
Related Publications: IbPRIA'2003, VLBV'2003
Some videos (mpeg): [Video 1], [Video 2]
SSD Based Tracking
SSD based tracking. Planar tracking can be used for face tracking. We have extended a well known framework for planar tracking (see [Hager98]) with a projective motion model. Additionally, using a calibrated camera, it is possible to estimate the 3D pose of the planar object. We have also developed a procedure to select the most informative pixels of the target template image for faster tracking.
Related Publications: ICPR'2002, ICIP'2002
Some videos (mpeg):  [Video 1], [Video 2]
Colour Based Tracking
Colour Based Tracking. Tracking using colour is difficult when sudden light colour changes take place. We have extented a well known colour constancy algorithm, Grey World, to deal with such situations. The result is more robust than widely used RGB-normalisation, although it is not perfect either.
Related Publications: SNRFAI'2001, CAIP'2001, IMP&COMM2001
Some videos (mpeg): [ light colour change], [RGB-normalised], [DGW].
SSD Based Tracking
Robust face tracking. All the algorithms based on a simple visual cues fail in some circumstances. The key idea is to use this simple algorithms together in order to get robustness. All this algorithms should be "orthogonal" in the sense of having different fail conditions.
Related Publications: CAEPIA'1999
Some videos (mpeg): Colour and SSD tracking