The problem of model selection --- automatically choosing the correct function to describe a data set --- is relevant to many areas of computer vision. Many model selection criteria have been used in the vision literature and many more have been proposed in statistics, but the relative strengths of these criteria have not been analyzed in vision. Using the problem of surface reconstruction as our context, we analyze existing criteria using simulations and real data, introduce new criteria from statistics, develop novel criteria capable of handling unknown error distributions and outliers, and extend model selection criteria to apply to the surface merging problem. The new surface merging rules improve upon previous results, and work well even at small step heights (h = 3oe) and crease discontinuities. Our results show that when the error distribution is known (at least approximately), Bayesian criteria for model selection and surface merging introduced here works best, although for tim...

An automatic word classification system has been designed which processes word unigram and bigram frequency statistics extracted from a corpus of natural language utterances. The system implements a type of simulated annealing which employs an average class mutual information metric. Resulting classifications are hierarchical, allowing variable class granularity. Words are represented as structural tags --- unique n-bit numbers the most significant bit-patterns of which incorporate class information. Therefore, access to a structural tag immediately provides access to all classification levels for the corresponding word. The classification system has successfully revealed some of the structure of two natural languages, from the phonemic to the semantic level. The system has been favourably compared --- directly and indirectly --- with other word classification systems. Class based interpolated language models have been constructed to exploit the extra information supplied by structural...

We present a review of some recently developed techniques in the field of natural language processing. This area has witnessed a confluence of approaches which are inspired by theories from linguistics and those which are inspired by theories from information theory: statistical language models are becoming more linguistically sophisticated and the models of language used by linguists are incorporating stochastic techniques to help resolve ambiguities. We include a discussion about the underlying similarities between some of these systems and mention two approaches to the evaluation of statistical language processing systems. 1 Introduction Within the last decade, a great deal of attention has been paid to techniques for processing large natural language copora. The purpose of much of this activity has been to refine computational models of language so that the performance of various technical applications can be improved (e.g. speech recognisers [67], speech synthesisers [32], optica...

A method is described for the in vitro culture of 17.5-day rat visceral yolk sac. Tissue survival was good as judged by light and electron microscopy. The rate of pinocytic uptake of ~25I-labeled polyvinylpyrrolidone by the tissue was constant both within and between experiments. Within the concentration range 0.15-24 ~tg/ml, the a2~I-labeled polyvinylpyrrolidone neither stimulated nor inhibited pinocytosis. The system offers many advantages in the quantitative study of the physical basis of pinocytosis. Many important questions concerning pinocytosis remain unanswered. For example, do pinocyticaily ingested solutes enter chiefly in free solution, or adsorbed to the plasma membrane? By what means and to what extent can pinocytosis be stimulated and inhibited? Such questions can be answered only by experiments in which pinocytosis

In this paper, we are concerned with the problem of deciding whether a fitted model accurately describes the data to which it has been fitted. We have developed an effective method of testing the lack-of-fit of a parametric model to data, with applications to the computer vision problems of robust estimation, model selection, and curve and surface segmentation. The benefits of this technique are high sensitivity (large response to small outliers) and very low dependence on the noise distribution of the input data. Our test is new to the computer vision community in several ways: ffl We look at the distribution of the residual errors, rather than basing statistics directly on their values. ffl We assume a broad enough class of distributions as to be essentially distribution independent. ffl The test requires no knowledge of the sensor noise level, and its response is essentially independent of that level. We present results of experiments that compare the test with the standard Ø 2...

The choice of shape representation and the extraction of such representations from images is one of the great challenges of computer vision. This thesis addresses these issues by examining a number of topics in curve representations. Beginning with an examination of the conic fitting problem, a new linear ellipse fitter is developed. Previous ellipse-specific methods have been computationally expensive, and previous linear methods have fitted general conics, rather than ellipses, to the data. The new algorithm is compared with several others and is shown to be extremely stable and insensitive to noise. The comparison is itself of interest as it focusses on the behaviour of the algorithms under occlusion rather than noise, demonstrating that this is the parameter to which they are most sensitive. A comprehensive evaluation of conic fitting algorithms then follows, concluding that occlusion sensitivity is one of the key characteristics of the conic fitting problem. This survey is in itself of interest as it provides specific recommendations for practitioners in the field. The second part of the thesis deals with the question of deciding how well a model describes a given set of data. Two new techniques are discussed, both of which are independent of the noise level of the data, and which are therefore applicable to a wide range of automated processes. The run-distribution test of Chapter 5 is an effective method of determining a posteriori whether a given model accurately describes a data set. Comparisons with a number of standard tests indicate that the run-distribution test outperforms them unless the true noise level is known. The sum-of-variance metric of Chapter 6, on the other hand, provides a parameter-free method of segmenting a dataset into piecewise smooth segments. The behaviour of the metric is demonstrated

The spatial multivariate median has a long history as an alternative to the sample mean. Its transformation-retransformation (TR) sample version is affine equivariant, highly robust, and computationally easy. More recently, an entire TR spatial multivariate quantile function has been developed and applied in practice along with related rank functions. However, as quantile levels move farther out, robustness of the TR sample version as measured by breakdown point decreases to zero, a serious limitation in applications such as outlier detection and setting inner 50%, 75%, and 90 % quantile regions. Here we introduce a new device, “spatial trimming”, and with it solve two problems of general scope and application: (i) the need for robustification of the TR sample spatial quantile function and its closely related depth, outlyingness, and rank functions, and (ii) the need for a computationally easy, robust, and affine equivariant scatter estimator. Improvements in robustness accomplished by spatial trimming are confirmed by improved breakdown points and illustrated using simulated and actual data. Other applications of spatial trimming are

ABSTRACT The elastimeter method was applied to the single muscle fiber of the frog semitendinosus to obtain the elastic moduli of the sarcolemma and myoplasm, as well as their relative contributions to resting fiber tension at different extensions. A bleb which was sucked into a flat-mouthed pipette at the fiber surface separated into an external sarcolemmal membrane and a thick inner myoplasmic region. Measurements showed that the sarcolemma does not contribute to intact fiber tension at sarcomere lengths below 3 p. It was estimated that the sarcolemma contributed on the order of 10 % to intact fiber tension at sarcomere lengths between 3 and 3.75 Ia, and more so with further extension. Between these sarcomere lengths, the sarcolemma can be linearly extended and has a longitudinal elastic modulus of 5 x 106 dyne/cm2 (assuming a thickness of 0.1 pu). Resistance to deformation of the inner bleb region is due to myoplasmic elasticity. The myoplasmic elastic modulus was estimated by use of a model and was used to predict a fiber length-tension curve which agreed approximately with observations.

In this paper we will use functional networks to model some linear and non linear relations among variables. In particular, our method allows us to discover adequate transformations of the response and/or the explanatory variables in multiple linear regression. If we apply this method to a heteroscedastic linear problem, we can estimate all the parameters involved in the model. Furthermore, we will tackle the estimation of classification functions. The proposed approach is compared with other statistical methods for classification and regression problems. Finally, the performance of the proposed procedure is illustrated by a simulation study and by real-life data sets.

Introduction and Disclaimer These worked examples illustrate the use of the BUGS language and sampler in a wide range of problems. They contain a number of useful "tricks", but are certainly not exhaustive of the models that may be analysed. We emphasise that all the results for these examples have been derived in the most naive way: in general a burn-in of 500 iterations and a single long run of 1000 iterations. This is not recommended as a general technique: no tests of convergence have been carried out, and traces of the estimates have not even been plotted. However, comparisons with published results have been made where possible. Times have been measured on a 60 MHz superSPARC: a 60 MHz Pentium PC appears to be about 4 times slower, and a 30 MHz superSPARC about 2 times slower. Users are warned to be extremely careful about assuming convergence, especially when using complex models including errors in variables, crossed random effects and intrinsi