Tim J. Dennis

Epipolar line estimation and rectification for stereo image pairs

The assumption that epipolar lines are parallel to image scan lines is made in many algorithms for stereo analysis. If valid, it enables the search for corresponding image features to be confined to one dimension and, hence, simplified. An algorithm that generates a vertically aligned stereo pair by warped resampling is described. The method uses grey scale image matching between the components of the stereo pair but confined to feature points.

An adaptive implementation of the SUSAN method for image edge and feature detection

Edge and feature points are basic low level primitives for image processing. The justification is that such points convey a lot of structural information about the surfaces of the objects captured in the imaging process. Because of this, edge and feature detection are two of the most common operations in image analysis. Despite the variety of operators available for these purposes, there are still some drawbacks common to most of them, even the more sophisticated ones. Among them, the problems of connectivity and shape of junctions by edge detectors, and the dichotomy between reliability of detection and localization accuracy for feature detectors are particularly outstanding. The present work proposes an adaptive implementation of the SUSAN method for edge and feature detection. The aim is the development of an algorithm that offers some performance improvement in tackling these limitations.

Video subband VQ coding at 64 kbit/s using short-kernel filter banks with an improved motion estimation technique

Abstract A motion compensated interframe prediction (MCIP) coder incorporating short-kernel subband filter-banks with an improved motion estimation technique is described. The coder is based on two-dimensional subband decomposition of the error image obtained after motion compensated interframe prediction. The luminance and chrominance subbands, generated after the decomposition process, form image vectors which are then quantized using vector quantization. The block matching method works well with low velocities and low noise content in the image frames. For a motion compensation estimation algorithm to cope with higher activity image frames, it is important that rapid movements, having high-amplitude motion vectors, are effectively recognized. It is also essential that the motion estimation performance does not degrade due to noise in the previously decoded frame. To achieve this, a combination of phase correlation and block search methods is applied sequentially to a hierarchy of low-pass filtered and subsampled images, while the motion vectors are progressively refined at each step of the process. Details of the proposed coding algorithm and simulation results operating at 64 kbits/s are presented.

Three-dimensional parameter estimation from stereo image sequences for model-based image coding

Abstract It is shown that the analysis of moving image sequences for 3D modelling can be performed in a relatively straightforward manner if the scene is captured in stereo. Output from a stereo disparity estimation process using calibrated cameras gives absolute 3D surface coordinates from a single stereo pair. When combined with monocular motion cues, the true 3D motion parameters of moving objects can be accurately calculated. Further analysis enables segmentation of body elements according to motion while the 3D surface feature structure, although available from the start, can be integrated and checked for anomalies over the sequence. These results are expected to alleviate the known problems of ambiguity suffered by monocular-source model-based coders.

Stereo disparity computation in the DCT domain using genetic algorithms

The objective of stereo matching is, given a point in one view of a scene, to find the homologous point in another view from an adjacent camera. We use the discrete cosine transform (DCT) with evolutionary methods as an approach to solve the correspondence problem. The matching process is implemented by generating for each image block sets of DCT coefficients, which when reverse transformed become the estimated disparity map for that block. Assumed statistical properties of the AC coefficients are taken into account, while the DC component is estimated from a preliminary block matching stage. Matching is performed in the image domain using an intensity similarity measure. The job of optimising the DCT coefficients, whose inverse transform gives the disparity, is carried out by a biologically inspired optimisation technique: the genetic algorithm.

Simulated annealing and iterated conditional modes with selective and confidence enhanced update schemes

Proposes a selective update scheme for both SA (simulated annealing) and ICMs (iterated conditional modes) which only visits sites within inhomogeneous neighborhoods. A second scheme is proposed to enhance the update confidence at each site by incorporating contextual information in terms of neighbor label class probability distributions, instead of their current realizations. The two update schemes reduce the computation demand and improve estimation accuracy. Both schemes are tested on a noise-contaminated Markov random field test image. The results show that ICM and SA with selective update achieve a computational savings of five times on average, without introducing noticeable degradation. The confidence enhanced update scheme, working with SA and ICM, much improves the final estimation accuracy. In particular for ICM, it produces similar results to those of SA, but uses only a fraction of the iterations needed by the latter.<<ETX>>

Efficient Representation of 3D Human Head Models.

Human head models are used in many applications of computer graphics such as computer animation and cyberspace communications. Spherical harmonic shape representation decomposes the surface into its spatial surface frequency components by 3D multiresolution techniques. This paper demonstrates that spherical harmonic representation is particularly suited for modelling of the general 3D human head shape due to the originally rounded surface geometry. Operating from range data the SH representation ignores fine surface detail and retains the general shape suitable for texture mapping. The SH surface is represented with a limited number of bases function coefficients, analogous to Fourier series decomposition. It is shown that the total data amount to represent any human head SH model can be limited to a few thousand bits, with quantized and variable-length coded parameters. The SH models can be reconstructed from the parameters in wire-frame format at the desired polygon resolution and post-processed with conventional graphics utilities.

A zero-crossing edge detector with improved localization and robustness to image brightness and contrast manipulations

To characterize edges, this work proposes a zero-crossing detector that attempts to provide a more precise localization of edges together with robustness to brightness and contrast variation in images. Since the algorithm aims to be insensitive to image histogram manipulations, it is based on zero-crossings of the second derivative, and not on first derivative measurements. In order to provide good localization while preserving the true shape of the edges, the detector is based on precise mathematical expressions that represent image derivatives as a vector field, instead of using the Laplacian as a scalar approximation for the image second derivatives. This way, the proposed detector can be viewed as a Marr-Hildreth operator with a more exact mathematical formulation. Finally, to avoid false responses and provide thin edges, some additional checking is performed over the edge candidate points. The method showed good results when compared with well-known edge detectors.

Statistical analyses of disparity maps and disparity compensated residuals in the presence of occlusions

The doubling of the data rate in stereo image systems causes problems for archiving and transmission, and compression is obviously desirable. In general, stereo codecs code one of the views as the reference using a conventional method like MPEG-2 assuming two basic sources, and the disparity map is calculated to aid the coding process. It can be coded using the DCT, another lossy method, or a lossless compression. The other view is either coded directly, or is restored from a combination of the reference view, the disparity data and a residual error image. In transform coding, inappropriate or coarse quantisation may cause coding artifacts. This work analyses and models the distribution of the 2D DCT coefficients of disparity maps and of disparity compensated residuals in the presence of occlusions, highlighting some points that should be considered when going from the monocular to the multiple-camera codec design process.

Evolutionary Methods Applied to Binocular Disparity Estimation

One problem on binocular disparity estimation is, given a point in one view of a scene, to find the homologous point in another view from an adjacent camera. In this work, we use the discrete cosine transform (DCT) with evolutionary methods as an approach to obtain the disparity map. The disparity estimation process is implemented by generating for each image block sets of DCT coefficients. The job of optimising the DCT coefficients, whose inverse transform gives the disparity map, is carried out by a optimisation technique inspired by natural evolution: the genetic algorithm. Matching is performed in the image domain using an intensity similarity measure.