Showing papers presented at "British Machine Vision Conference in 1992"

Active shape models - 'Smart Snakes'.

Abstract: We describe ‘Active Shape Models’ which iteratively adapt to refine estimates of the pose, scale and shape of models of image objects. The method uses flexible models derived from sets of training examples. These models, known as Point Distribution Models, represent objects as sets of labelled points. An initial estimate of the location of the model points in an image is improved by attempting to move each point to a better position nearby. Adjustments to the pose variables and shape parameters are calculated. Limits are placed on the shape parameters ensuring that the example can only deform into shapes conforming to global constraints imposed by the training set. An iterative procedure deforms the model example to find the best fit to the image object. Results of applying the method are described. The technique is shown to be a powerful method for refining estimates of object shape and location.

Training Models of Shape from Sets of Examples

Abstract: A method for building flexible shape models is presented in which a shape is represented by a set of labelled points. The technique determines the statistics of the points over a collection of example shapes. The mean positions of the points give an average shape and a number of modes of variation are determined describing the main ways in which the example shapes tend to deform from the average. In this way allowed variation in shape can be included in the model. The method produces a compact flexible ‘Point Distribution Model’ with a small number of linearly independent parameters, which can be used during image search. We demonstrate the application of the Point Distribution Model in describing two classes of shapes.

Feature-based correspondence: an eigenvector approach

Abstract: We propose a novel method for performing point-feature correspondence based on a modal shape description. Introducing shape information into a low-level matching process allows our algorithm to cope easily with rotations and translations in the image plane, yet still give a dense correspondence. We also show positive results for scale changes and small skews, and demonstrate how reflectional symmetries can be detected.

Trainable method of parametric shape description

Abstract: We have developed a trainable method of shape representation which can automatically capture the invariant properties of a class of shapes and provide a compact parametric description of variability. We have applied the method to a family of flexible ribbons (worms) and to heart shapes in echocardiograms. We show that in both cases a natural parameterisation of shape results.

Classification of breast tissue by texture analysis

Abstract: The identification of glandular tissue in breast X-rays (mammograms) is important both in assessing asymmetry between left and right breasts, and in estimating the radiation risk associated with mammographic screening. The appearance of glandular tissue in mammograms is highly variable, ranging from sparse streaks to dense blobs. Fatty regions are generally smooth and dark. Texture analysis provides a flexible approach to discriminating between glandular and fatty regions. We have performed a series of experiments investigating the use of granulometry and texture energy to classify breast tissue. Results of automatic classifications have been compared with a consensus annotation provided by two expert breast radiologists. On a set of 40 mammograms, a correct classification rate of 80% has been achieved using texture energy analysis.

Model-based image interpretation using genetic algorithms

Abstract: We describe the application of genetic algorithms in model-based image interpretation. The delineation of left ventricular boundaries in apical 4-ehamber echocardiograms is used as an illustrative exemplar. The suitability of genetic algorithms for the model/objective-function/search procedure is presented.

Performance Characterization in Computer Vision.

Abstract: Computer vision algorithms are composed of different sub-algorithms often applied in sequence. Determination of the performance of a total computer vision algorithm is possible if the performance of each of the sub-algorithm constituents is given. The problem, however, is that for most published algorithms, there is no performance characterization which has been established in the research literature. This is an awful state of affairs for the engineers whose job it is to design and build image analysis or machine vision systems.

Face Recognition by Computer

Abstract: We describe a coding scheme to index face images for subsequent retrieval, which seems effective, under some conditions, at coding the faces themselves, rather than particular face images, and uses typically 100 bytes. We report tests searching a pool of 100 faces, using as cue a different image of a face in the pool, taken 10 years later. In two of three tests with different faces, the target face best matches the corresponding cue.

Ellipse detection and matching with uncertainty

Abstract: This paper considers the fitting of ellipses to edge-based pixel data in digital images. It infers the ellipses as projected circles, viewed obliquely, and uses this information to constrain the position, viewpoint and scale of model objects projected onto the image. A central theme of the work is the explicit use of uncertainty, which reflects observation errors in the image data, and which is propagated through the model matching stages, providing a consistent and robust representation of errors.

A Matching and Tracking Strategy for Independently Moving Objects

Abstract: We present a robust and inherently parallel strategy for tracking “corner” features on independently moving (and possibly non-rigid) objects. The system operates over long, monocular image sequences and comprises two main parts. A matcher performs two-frame correspondence based on spatial proximity and similarity in local image structure, while a tracker maintains an image trajectory (and predictor) for every feature. The use of low-level features ensures an opportunistic and widely applicable algorithm. Moreover, the system copes with noisy data, predictor failure, and occlusion and disocclusion of scene structure. Motion and scene analysis modules can then be built onto this framework. The algorithm is aimed at applications with small inter-frame motion, such as videoconferencing.

Camera Calibration using Vanishing Points

Abstract: This paper describes a method for measuring the intrinsic parameters of a camera. If aspect ratio is already known, the method requires two images of a plane; if aspect ratio is not known, the method requires three images of the plane. (Equivalently, one image of n planes could be used instead of n images of one plane). The plane must be marked with four points in known configuration.

A New Class of Corner Finder

Abstract: An accurate, stable and very fast corner finder (for feature based vision) has been developed, based on a novel definition of corners, using no image derivatives. This note describes the algorithm and the results obtained from its use.

A Generic System for Image Interpretation Using Flexible Templates

Abstract: We describe a generic approach to image interpretation, based on combining a general method of building flexible template models with Genetic Algorithm (GA) search The method can be applied to a given image interpretation problem simply by training a Point Distribution Model (PDM), using a set of examples of the image structure to be located A local optimisation technique, developed for use with PDMs, has been incorporated into the GA search with the aim of improving the speed of convergence and optimality of solution We present results, from three practical applications, demonstrating that the new method offers significant improvements when compared to previously reported approaches to flexible template matching The benefits include the ability to deal with different domains of application using a standard method, the ability to deal with complex multi-part models and improved search performance

Affine and Projective Structure from Motion

Abstract: We demonstrate the recovery of 3D structure from multiple images, without attempting to determine the motion between views. The structure is recovered up to a transformation by a 3D linear group — the affine and projective group. The recovery does not require knowledge of camera intrinsic parameters or camera motion.

On Local Matching of Free-Form Curves

Abstract: Geometric matching in general is a difficult unsolved problem in computer vision. Fortunately, in many practical applications, some a priori knowledge exists which considerably simplifies the problem. In visual navigation, for example, the motion between successive positions is usually either small or approximately known, but a more precise registration is required for environment modeling. The algorithm described in this paper meets this need. Objects are represented by free-form curves, i.e., arbitrary space curves of the type found in practice. A curve is available in the form of a set of chained points. The proposed algorithm is based on iterativeiy matching points on one curve to the closest points on the other. A least-squares technique is used to estimate 3-D motion from the point correspondences, which reduces the average distance between curves in the two sets. Both synthetic and real data have been used to test the algorithm, and the results show that it is efficient and robust, and yields an accurate motion estimate.

Stereo Without Disparity Gradient Smoothing: a Bayesian Sensor Fusion Solution

Abstract: A maximum likelihood stereo algorithm is presented that avoids the need for smoothing based on disparity gradients, provided that the common uniqueness and monotonic ordering constraints are applied. A dynamic programming algorithm allows matching of the two epipolar lines of length N and M respectively in O(N M) time and in O(N) time if a disparity limit is set. The stereo algorithm is independent of the matching primitives. A high percentage of correct matches and little smearing of depth discontinuities is obtained based on matching individual pixel intensities. Because feature extraction and windowing are unnecessary, a very fast implementation is possible.

Invariants of a pair of conics revisited

Abstract: The invariants of a pair of quadratic forms and a pair of coplanar conics are revisited following [1, 5, 2]. For a given pair of conies, with their associated matrices C1 and C2, we show that Trace (C-12C1), Trace (C-11C2) and ∣C1∣/∣C1∣ are only invariants of associated quadratic forms, but not invariants of the conies. Two of true invariants of the conies are \(\frac{{Trace(C_2^{ - 1} C_1 )}}{{(Trace(C_2^{ - 1} C_1 ))^2 }}\frac{{\left| {C_2 } \right|}}{{\left| {C_1 } \right|}}and\frac{{Trace(C_2^{ - 1} C_1 )}}{{(Trace(C_2^{ - 1} C_1 ))^2 }}\frac{{\left| {C_1 } \right|}}{{\left| {C_2 } \right|}}\). Then the true invariants of the conies are geometrically interpreted, in terms of cross ratios, through the common self-polar triangle of the two conies.

Segmentation of MR images using neural nets

Abstract: A two tier system for MR image segmentation is proposed. The first stage involves probabilistically classifying the pixels of the input image based on second order grey level statistics obtained from the co-occurrence matrix. These probabilities then form the input to a multi-layer perceptron, allowing contextual constraints to be applied to local areas of the image thereby improving the spatial coherence of the classification.

Multiresolution Estimation of 2-d Disparity Using a Frequency Domain Approach

Abstract: An efficient algorithm for the estimation of the 2-d disparity between a pair of stereo images is presented. Phase based methods are extended to the case of 2-d disparities and shown to correspond to computing local correlation fields. These are derived at multiple scales via the frequency domain and a coarse-to-fine 'focusing' strategy determines the final disparity estimate. Fast implementation is achieved by using a generalised form of wavelet transform, the multiresolution Fourier transform (MFT), which enables efficient calculation of the local correlations. Results from initial experiments on random noise stereo pairs containing both 1-d and 2-d disparities, illustrate the potential of the approach.

Using colour templates for target identification and tracking

Abstract: A surveillance system is presented which uses colour cues to track people moving in sparse crowd scenes. The erratic motion of the targets, together with their changeable outline, means that they are conventionally difficult to model. However, by examining the colours present in an object in a given frame of a sequence, and looking for these in a later frame, identification and tracking are achieved. The colours are transformed into a template space in which it is easy to match objects to those held in a database, so that even when a person is occluded or disappears totally from view they may be re-located as soon as they can be clearly seen again. A hierarchical approach to template storage and searching reduces the effort required to search through the database and ensures that the system is efficient even with a database containing the templates of hundreds of people. Since work is carried out on a moving sequence, the problems of changing object shape and maintaining colour constancy are minimised by allowing for small changes in these parameters between frames and by continually updating each object’s template.

Linear Algorithms for Object Pose Estimation

Abstract: This paper concerns the estimation of object pose in scenes where objects are located on the ground plane which has known orientation and position w.r.t. the camera. Novel algorithms are described, based on the concept of interpretation planes and that of pencils of planes. The methods are linear, computationally simple, and give unique and closed-form solutions, thus eliminating many of the problems associated with the existing pose recovery algorithms. They require a minimum of two 2D-3D line correspondences. Experimental results are included which show that the proposed algorithms are robust to noise, and capable of accurate pose recovery using real images of outdoor scenes.

Machine Vision Inspection of Web Textile Fabric

Abstract: The paper describes instrumentation which uses machine vision to inspect rolls of web textile fabric in real time. This involves detection of “message” signals arising from defects buried in noise caused by fabric structure. Analogy with the detection of targets in radar and sonar is exploited to provide effective signal processing. The hardware implementation achieves efficiency with economy by using standard devices wherever possible — such as a CCD linescan camera for sensing and a 486 PC as host processor. A special interface card is provided which compensates for deterministic noise and eliminates more than 99% of the redundant data gathered by the camera.

Coarse Image Motion for Saccade Control

Abstract: We describe a 2D vision module that estimates the motion of moving objects for the purpose of driving saccadic head/eye motions to fixate them. Robustness and a fast reaction time are the main requirements of the module. The apparently moving background is segmented from the moving objects in the scene using a prediction of the background flow obtained from head odometry. Subsequently the velocities of the detected objects are determined using a least-square method to solve the aperture problem. The algorithm has been implemented at frame rate (25Hz) on a network of five transputers, and has a latency of approximately 0.06s.

From Features to Perceptual Categories

Abstract: First we review an analysis of conditions that should be met if features are to provide robust inferences about world properties. Features meeting these conditions provide indices into especially useful categories of visual properties. Then we show that for a given set of elemental concepts the categories associated with these properties have a natural hierarchical (specialization) structure. We argue that this structure provides constraints on the form and type of categories that are inferred when visual objects are classified.

Statistical Analysis of a Stereo Matching Algorithm

Abstract: This paper discusses the problems of image processing algorithm design and comparison and suggests that a suitable approach may be to model algorithms. We introduce the corner matching algorithm which we have used to provide reliable data for 3D computation modules [5][6]. The development of a simple model of the matching process permits the understanding of the influence of various parameters in the matching algorithm. This model also allows optimisation of the algorithm using data distributions obtained from representative scenes.

Active Contours using Finite Elements to Control Local Scale

Abstract: Finite elements allow smoothness to be enforced on the measurement of the image-dependent term in active contours. This improves the stability of the solution, with less computational cost than is incurred by increasing the number of elements. Performance is best when the size of the element matches the scale of the image detail sought.

A method of obtaining the relative positions of four points from three perspective projections

Abstract: According to Ullman’s Structure-from-Motion Theorem [U79], three orthogonal projections of four points in a rigid non-planar configuration uniquely determine their structure, and the relative orientations of the three views, up to a reflection in the image plane. It is here shown that a corresponding result holds for the more general “para-perspective” case, and leads to a rapidly convergent algorithm for the fully perspective case. Unless the four points are nearly coplanar, or the images closely similar, the output of this algorithm is not unduly sensitive to errors in the image coordinates.

Estimation of Cloud Cover using Colour and Texture

Abstract: We describe methods for using colour and texture to discriminate cloud and sky in images captured using a ground based colour camera. Neither method alone has proved sufficient to distinguish between different types of cloud, and between cloud and sky in general. Classification can be improved by combining the features using a Bayesian scheme.

Multistage Combined Ellipse and Line Detection

Abstract: This paper describes an algorithm for the detection of ellipses and lines in image edge data. Connected edge pixels are transformed into polygonal approximations by a two stage algorithm. Then a second two stage algorithm replaces combinations of lines by ellipses if the ellipse fit is better. For each algorithm a combination of splitting and merging of the data is used to enable global and local constraints to fit different representations to the pixel data. All merging decisions use a significance measure to replace a number of representations by a single representation which removes the need for thresholds in the algorithms. The structure of the algorithm allows any particular representation to describe the data e.g. parabolae, splines, etc. instead of ellipses.