Search

We present a new method for automatic landmark extraction from the contours of biological specimens. Our ultimate goal is to enable automatic identification of biological specimens in photographs and drawings held in a database. We propose to use active appearance models for visual indexing of both photographs and drawings. Automatic landmark extraction will assist us in building the models. We describe the results of using our method on drawings and photographs of examples of diatoms, and present an active shape model built using automatically extracted data.

We present a system capable of producing video-realistic videos of a speaker given audio only. The audio input signal requires no phonetic labelling and is speaker independent. The system requires only a small training set of video to achieve convincing realistic facial synthesis. The system leams the natural mouth and face dynamics of a speaker to allow new facial poses, unseen in the training video, to be synthesised. To achieve this we have developed a novel approach which utilises a hierarchical and non-linear PCA model which couples speech and appearance. We show that the model is capable of synthesising videos of a speaker using new audio segments from both previously heard and unheard speakers. The model is highly compact making it suitable for a wide range of real-time applications in multimedia and telecommunications using standard hardware.

An alternative to representing curves at a single scale or a fixed number of multiple scales is to represent them only at their natural (i.e. most significant) scales. This allows all the important information concerning the different sized structures contained in the curve to be explicitly represented without the overhead of redundant representations of the curve. This paper describes several approaches to determining the local natural scales of curves. That is, various possibly overlapping sections of the curve should be represented at certain scales depending on their shape. The merits and drawbacks of the techniques are described, and the results of implementing one of them are shown.

Ellipses seen in an image may be the 2-D projection of 3-D circles from the scene. Given this assumption, ellipses can be grouped into perceptual groups from which inferences about the 3-D structure of objects can be made. Methods are proposed for extracting groupings corresponding to surfaces of revolution. A Hough transform approach is used for grouping, after which the confidence in the plausibility of the perceptual group is improved by detecting symmetry groupings. >

A method is described for segmenting edge data into a combination of straight lines and elliptic arcs. The two-stage process first segments the data into straight line segments. Ellipses are then fitted to the line data. This is much faster than curve fitting directly to pixel data since the lines provide a great reduction in data. Segmentation is performed in the paradigm suggested by D.G. Lowe (1987). A measure of significance is defined that produces a scale-invariant description and allows the replacement of sequences of line segments by ellipses without requiring any thresholds. A method for fitting ellipses to arbitrary curves, essential for this algorithm, has been developed, based on an iterative Kalman filter. This is guaranteed to produce an elliptical fit even though the best conic fit may be a hyperbola or parabola. >

The authors study neural network overfitting on synthetically generated and real remote sensing data. The effect of overfitting is shown by: 1) visualising the shape of the decision boundaries in feature space during the learning process, and 2) by plotting the classification accuracy of independent test sets versus the number of training cycles. A solution to the overfitting problem is proposed that, involves pre-processing the training data. The method relies on obtaining an increase of spectral coherence of individual training classes by applying k-nearest neighbour filtering. Points in feature space with class labels inconsistent with those of the majority of their neighbours are removed. This effectively simplifies the training data, and removes outliers and local inconsistencies. It is shown that using this approach can reduce the overfitting effect and increase the resulting classification accuracy.

The proliferation of content based image retrieval techniques has highlighted the need to understand the relationship between image clustering based on low-level image features and image clustering made by human users. In conventional image retrieval systems, images are typically characterized by a range of features such as color, texture, and shape. However, little is known to what extent these low-level features can be effectively combined with information visualization techniques such that users may explore images in a digital library according to visual similarities. The authors compare and analyze a number of Pathfinder networks of images generated based on such features. Salient structures of images are visualized according to features extracted from color, texture, and shape orientation. Implications for visualizing and constructing hypermedia systems are discussed.

" #$ %&'" % '" $ ( '" ) * + * + , ( *

Detecting landslides and monitoring their activity is of great relevance for natural hazard assessment and disaster prevention in hilly areas. Very high resolution optical satellite data is now becoming available, and to evaluate the potential application of these new images to detect ground surface changes as a result of landsliding, in previous work we developed and applied change detection and thresholding methods on digital aerial photographs over the Tessina landslide in Veneto, Italy . This is a nearly 3-km long complex landslide consisting of rotational slides in its head and a mudflow in the remaining, largest part of the body. The landslide has developed in Eocene flysch materials partly covered by colluvial and glacial till sediments. The landslide was first triggered in 1960 and has since undergone a number of reactivations. The physical setting of the area and landslide dynamics are further described in . The landslide currently affects an area including a number of small villages and mixed woodland-grassland landcover. It generally appears as a distinct bright feature on visible-wavelength panchromatic images because of outcropping soil under disrupted vegetation (figure 1a&b). This paper focuses on applying genetic programming techniques to detect and monitor landslides from optical remote sensing data. To this end, we have used radiometrically mormalised and orthorectified multitemporal aerial photographs at 1m ground resolution over the Tessina landslide area .

The main purpose of this research is to develop methods for automatic identification of biological specimens in digital photographs and drawings held in a database. Incorporation of taxonomic drawings into a visual indexing system has not been attempted to date. Diatoms are a single cell microscopic algae that provide a particularly suitable case study. Identification of diatoms is a challenging task due to the huge number of the species, blurred boundaries between species, and life cycle related shape changes. A novel model based on principal curves representing the life cycle related shape variation of a number of diatom species has been developed. Our model is suitable for reconstruction purposes, allowing us to produce drawings of a variety of diatom shapes, thus providing a link between the photographs and drawings. We present the classification results of photographed and drawn specimens based on the model and compare our results to another recent system for diatom identification. Finally, given a diatom specimen, we are able not only to identify the species it belongs to but also to pinpoint the stage in the life cycle it represents.

Convexity estimators are commonly used in the analysis of shape. In this paper we define and evaluate a new easily computable measure of convexity for polygons. Let be an arbitrary polygon. If denotes the perimeter in the sense of metrics of the polygon obtained by the rotation of by angle with the origin as the center of the applied rotation, and if is the Euclidean perimeter of the minimal rectangle having the edges parallel to coordinate axes which includes such a rotated polygon , then we show that defined as can be used as an estimate for the convexity of .S everal desirable properties of are proved, as well.

In this paper we define a function R(P) which is defined for any polygon P and which maps a given polygon P into a number from the interval (0, 1]. The number R(P) can be used as an estimate of the rectilinearity of P. The mapping R(P) has the following desirable properties: - any polygon P has the estimated rectilinearity R(P) which is a number from (0, 1]; - R(P)=1 if and only if P is a rectilinear polygon, i.e., all interior angles of P belong to the set {?/2, 3?/2}; - inf P?? R(P) = 0, where ? denotes the set of all polygons; - a polygon's rectilinearity measure is invariant under similarity transformations. A simple procedure for computing R(P) for a given polygon P is described as well.

Colour histograms have been shown to be very effective in representing images for content-based retrieval systems. They are compact, robust and amenable to computer analysis. However, such histograms convey only global image properties and do not embody local colour information which is so important when comparing and contrasting images. We present a novel image coding scheme which captures some of this locally correlated colour information and improves the selectivity of the retrieval mechanism-an important issue for very large databases. The technique uses a histogram of features which represent local combinations of colour artifacts within an image, called colour n-grams. We outline the thrust of our approach and discuss the factors affecting the efficacy of the retrieval mechanism using a database of colour images. (6 pages)

Various methods for automatic change detection in multi-temporal LANDSAT-TM images are described. In contrast to most previous work in change detection, which has operated at a pixel level, we operate at a parcel level (within a minimum size of 25 Ha). This makes it easier to employ structural measures (e.g. based on edges, corners, and texture) as well as correlation methods since these approaches cannot be calculated at each pixel independently. A neural network is trained to combine the different change measures in an appropriate manner.

This paper describes a method for post-processing classified images to enable generalisation to be performed whilst maintaining or improving the accuracy of region boundaries. This is achieved by performing region growing, and incorporates both spatial context and spectral information. In contrast, few classifiers use any spatial context, and many post-processing techniques, such as iterative majority filtering, discard all spectral information. If class models are available these can also be included in the region growing process, otherwise, the algorithm operates in a data-driven mode, and locally estimates models for each region.1

We present a comparative study of the effects of applying pre-processing and post-processing to remote sensing data both in the spatial image domain and the feature domain. We use a neural network for classification since it is not biased by a priori assumptions about the distributions of the spectral values of the classes. Spatial smoothing was applied both as pre- and post-processing steps. Pre-processing involved smoothing the image spectral values by means of anisotropic diffusion, whereas iterative majority filtering was applied as a post- processing step to improve spatial coherence by reclassifying pixels. While it is common practice to filter the image before classification (smoothing) or after classification (iterative majority filtering) it is less obvious what happens if pre-processing is applied to the training or image data in feature space. To minimize the effect of noisy training pixels we applied a k- nearest neighbor filtering algorithm to the training data. This involved reclassifying each training pixel by the majority class of the set of k closest training pixels (in terms of Euclidean distance) in feature space. The procedure eliminates isolated training pixels and tends to produce more compact class clusters. The effects of all spatial and spectral filtering methods were validated by applying them to three different testcases.© (1994) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

A number of objects in the real world can be described as surfaces of revolution. These are a particular type of generalised cylinder with a straight axis whose 3D shape is formed by rotating a 2D plane about the axis. Examples of such objects are vases, many chess pieces, light bulbs, table lamps etc. This paper describes a number of techniques that can be used to recognise this class of object in a typical cluttered scene under perspective projection. Use is made of the symmetry of the occluding boundary, perceptual grouping of ellipses, 3D modelsand the hypothesis that an ellipse is a circle in the real world. 1. INTRODUCTION A large number of objects in the real world can be described as surfaces of revolution (SORs) e.g. vases, manychess pieces, light bulbs, table lamps, etc. These are a particular form of generalised cylinder, i.e. right, circular,straight, homogeneous generalised cylinders (RCSHGC's)15. SORs are an interesting type of object to investigatefor recognition as they contain straight and curved edges and surfaces and exhibit symmetry. In addition they arerelatively easy to describe as 3D models. Figure 1 shows views of the models used.

Lowe demonstrated a method for automatically segmenting and smoothing image curves by varying degrees. It was intended to remove noise and unnecessary fine detail, aiding subsequent processing such as grouping and matching. An alternative technique is described in this paper that is based on recursively subdividing the curve into alternative sets of sections. Rather than use thresholds on the values of curvature and its derivatives to determine the segmentation and degree of smoothing our technique is driven by three qualitative measures: (1) a criterion for selecting potential breakpoints, (2) a criterion for determining the amount of smoothing for curve sections, and (3) a significance measure that determines which sections form the best selection. The advantages of the technique are robustness, scale invariance, and the absence of parameters.© (1993) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

The distance transform has been proposed for use in computer vision for a number of applications such as matching and skeletonisation. This paper proposes two things: (1) a multi-scale distance transform to overcome the need to choose edge thresholds and scale and (2) the addition of various saliency factors such as edge strength, length and curvature to the basic distance transform to improve its effectiveness. Results are presented for applications of matching and snake fitting.

A novel statistical model based on texture and shape for fully automatic intraretinal layer segmentation of normal retinal tomograms obtained by a commercial 800nm optical coherence tomography (OCT) system is developed. While existing algorithms often fail dramatically due to strong speckle noise, non-optimal imaging conditions, shadows and other artefacts, the novel algorithm’s accuracy only slowly deteriorates when progressively increasing segmentation task difficulty. Evaluation against a large set of manual segmentations shows unprecedented robustness, even in the presence of additional strong speckle noise, with dynamic range tested down to 12dB, enabling segmentation of almost all intraretinal layers in cases previously inaccessible to the existing algorithms. For the first time, an error measure is computed from a large, representative manually segmented data set (466 B-scans from 17 eyes, segmented twice by different operators) and compared to the automatic segmentation with a difference of only 2.6% against the inter-observer variability.

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.

The authors describe a method to re-map animation parameters between multiple types of facial model for performance-driven animation. A facial performance can be analysed in terms of a set of facial action parameter trajectories using a modified appearance model with modes of variation encoding specific facial actions which can be pre-defined. These parameters can then be used to animate other modified appearance models or 3D morph-target-based facial models. Thus, the animation parameters analysed from the video performance may be re-used to animate multiple types of facial model. The authors demonstrate the effectiveness of the proposed approach by measuring its ability to successfully extract action parameters from performances and by displaying frames from example animations. The authors also demonstrate its potential use in fully automatic performance-driven animation applications.

The world (both natural and man-made) is constrained by physical laws to a finite number of basic patterns. Certain three dimensional relationships between features in a scene give rise to two dimensional relationships that are relatively viewpoint invariant and can be distinguished as non-accidental. Thus, those regularities observable in the image arise not by accident, but are projections of real regularities in the scene. The Gestalt movement proposed that the human visual system applied the principles of perceptual organisation to group image features manifesting those regularities. These were used to recover three dimensional structure from two dimensional views, providing cues to models for scene interpretation.

In order to improve the effectiveness of shape-based classification, there is an ongoing interest in creating new shape descriptors or creating new measures for descriptors that are already defined and used in shape classification tasks. Convexity is one of the most widely used shape descriptors and also one of the most studied in the literature. There are already several defined convexity measures. The most standard one comes from the comparison between a given shape and its convex hull. There are also some non-trivial approaches. Here, we define a new measure for shape convexity. It incorporates both area-based and boundary-based information, and in accordance with this it is more sensitive to boundary defects than exclusively area-based convexity measures. The new measure has several desirable properties and it is invariant under similarity transformations. When compared with convexity measures that trivially follow from the comparison between a measured shape and its convex hull then the new convexity measure also shows some advantages, particularly for shapes with holes.

Shape descriptors are used in many computer vision tasks. Convexity is one of the most widely used shape descriptors in practical applications and also one of the most studied in the literature. There are already several defined convexity measures. The most standard one comes from the comparison between a given shape and its convex hull, but there also other approaches. Independently of whether the convexity descriptors are area based or boundary based, all of them assume that the shape (or shape boundary) is completely known and that the measures apply to the complete data. A convexity measure is defined that can be applied to shapes with partially extracted boundaries. More formally, the new convexity measure deals with planar curves or with disconnected curves consisting of several planar curve segments.

A system capable of producing near video-realistic animation of a speaker given only speech inputs is presented. The audio input is a continuous speech signal, requires no phonetic labelling and is speaker-independent. The system requires only a short video training corpus of a subject speaking a list of viseme-targeted words in order to achieve convincing realistic facial synthesis. The system learns the natural mouth and face dynamics of a speaker to allow new facial poses, unseen in the training video, to be synthesised. To achieve this the authors have developed a novel approach which utilises a hierarchical and nonlinear principal components analysis (PCA) model which couples speech and appearance. Animation of different facial areas, defined by the hierarchy, is performed separately and merged in post-processing using an algorithm which combines texture and shape PCA data. It is shown that the model is capable of synthesising videos of a speaker using new audio segments from both previously heard and unheard speakers.

A method of segmenting curves into series of superelliptical arcs is presented. A superellipse is the 2-dimensional form of the superquadratic and can describe circles, ellipses, crosses, parallelograms and rounded rectangles with the same number of parameters. The superellipses are fitted using Powell's technique to minimise an appropriate error metric. A tree is used to represent a number of interpretations and the concept of significance used to choose the most perceptually correct description. Results show that perceptually good features are chosen to represent the various shapes that occur in images.

The retrieval of non-rigid 3D shapes is an important task. A common technique is to simplify this problem to a rigid shape retrieval task by producing a bending invariant canonical form for each shape in the dataset to be searched. It is common for these techniques to attempt to ``unbend'' a shape by applying multidimensional scaling to the distances between points on the mesh, but this leads to unwanted local shape distortions. We instead perform the unbending on the skeleton of the mesh, and use this to drive the deformation of the mesh itself. This leads to a computational speed-up and less distortions of the local details of the shape. We compare our method against other canonical forms and our experiments show that our method achieves state-of-the-art retrieval accuracy in a recent canonical forms benchmark, and only a small drop in retrieval accuracy over state-of-the-art in a second recent benchmark, while being significantly faster.

Digital images and videos have an increasingly important role in today's telecommunication and our everyday life in modern information society. In this paper, we explore the synergy between content-based image retrieval (CBIR) techniques and Pathfinder networks. Salient image features, based on colour, texture, edge orientation, and edge distance, are extracted. The structural modelling capabilities of Pathfinder are then applied to simplify and visualise the strongest interrelationships in the image database.

In this paper we present a novel automatic background substitution approach for live video. The objective of background substitution is to extract the foreground from the input video and then combine it with a new background. In this paper, we use a color line model to improve the Gaussian mixture model in the background cut method to obtain a binary foreground segmentation result that is less sensitive to brightness differences. Based on the high quality binary segmentation results, we can automatically create a reliable trimap for alpha matting to refine the segmentation boundary. To make the composition result more realistic, an automatic foreground color adjustment step is added to make the foreground look consistent with the new background. Compared to previous approaches, our method can produce higher quality binary segmentation results, and to the best of our knowledge, this is the first time such an automatic and integrated background substitution system has been proposed which can run in real time, which makes it practical for everyday applications.

http://www.bmva.org/bmvc/2002/papers/22/full_22.pdf; International audience; In this paper, we consider structure and motion recovery for scenes consisting of static and dynamic features. More particularly, we consider a single moving uncalibrated camera observing a scene consisting of points moving along straight lines converging to a unique point and lying on a motion plane. This scenario may describe a roadway observed by a moving camera whose motion is unknown. We show that there exist matching tensors similar to fundamental matrices. We derive the link between dynamic and static structure and motion and show how the equation of the motion plane (or equivalently the plane homographies it induces between images) may be recovered from dynamic features only. Experimental results on real images are provided, in particular on a 60frames video sequence.

International audience; One of the most popular methods to extract useful informations from an image sequence is the template matching approach. In this well known method the tracking of a certain feature or target over time is based on the comparison of the content of each image with a sample template. In this article, we propose an efficient robust template matching algorithm that is able to track targets in real time. Special attention is paid to occlusions handling and illumination variations.

International audience; This paper presents a real-time tracking technique for constructing spherical image mosaics from sequences of images. The main contribution is a tracking algorithm able to estimate in real-time camera rotations during the sequence, assuming that the optical center of the camera remain fixed. This assumption allows us to consider a sphere centered at this particular point and having its radius equal to the focal length of the camera. The problem is to estimate for each image of the sequence the position of the image plane relatively to this sphere and to back project the image to create the global spherical mosaic image. The proposed algorithm allows, during a first stage, to track in real time (less than 12ms for each image) the camera rotations around the optical center. This is done by using a real-time template matching algorithm, using the formalism described in [2], adapted in order to estimate only 3D camera rotations. After this stage, each one of the images of the sequence are located relatively to the spherical image. In a second step, the spherical image is build by back projecting on the sphere the images of the sequence. This paper is mainly devoted to the description of the first stage. Experimental results proving the efficiency of the proposed approach are given.