This study investigates the role F0 plays in the perception of prominence in American English. Raw, log and locally normalized measures of F0 were extracted from words in a 35K word corpus of spontaneous speech. Linear regression analyses were conducted to test the strength of these measures as cues to prominence, with prominence based on judgments made by ordinary listeners in real-time auditory perception. The Bayesian Information Criterion was used to further investigate whether these F0 measures cue prominence in a linear or piecewise linear function, corresponding to a linguistic model of prominence as a gradient or discrete feature. The results of this study show that F0 measures are similar to intensity measures in both their strength as cues to perceived prominence, and in signaling a discrete prominence distinction that distinguishes nonor weaklyprominent words from words with greater prominence. Our finding that F0 and intensity cue a discrete prominence distinction is compared with our prior finding that duration and word frequency signal gradient prominence distinctions. This apparent discrepancy is discussed in terms of the dual nature of prominence in English, as an expression of layered metrical (stress) structure in phonology, and as an expression of pragmatic focus.

This study investigated the relation between various acoustic features and prominence. Past research has suggested that duration, pitch, and intensity all play a role in the perception of prominence. In our past work, we found a correlation between these acoustic features and speaker agreement over the placement of prominence. The current study was motivated by a need to enrich our understanding of this correlation. Using the Bayesian information criterion, we show that the best model for a feature that cues prosody is not necessarily a single Gaussian. Rather, the best model depends on the feature. This finding has consequences for our understanding of the role of these features in the perception of prosody and for prosody recognition systems.

This paper describes a representation for people and animals, called a body plan, which is adapted to segmentation and to recognition in complex environments. The representation is an organized collection of grouping hints obtained from a combination of constraints on color and texture and constraints on geometric properties such as the structure of individual parts and the relationships between parts. Body plans can be learned from image data, using established statistical learning techniques. The approach is illustrated with two examples of programs that successfully use body plans for recognition: one example involves determining whether a picture contains a scantily clad human, using a body plan built by hand; the other involves determining whether a picture contains a horse, using a body plan learned from image data. In both cases, the system demonstrates excellent performance on large, uncontrolled test sets and very large and diverse control sets.

This paper describes a new representation for people and animals, called a body plan. The representation is an organized collection of grouping hints obtained from constraints on color, texture, shape, and geometrical relations. Body plans can be learned from image data, using established statistical learning techniques. Body plans are well adapted to segmentation and recognition in complex environments, such as the huge libraries of digitized images now becoming widely available. Two specific applications of body plans are presented: an algorithm that determines whether an image depicts a scantily clad human and an algorithm that learns and uses a body plan to find pictures of horses. Both algorithms demonstrate excellent performance on large, poorly controlled input data.

Retrieving images from very large collections, using image content as a key, is becoming an important problem. Finding objects in image databases is a big challenge in the field. The paper describes our approach to object recognition, which is distinguished by: a rich involvement of early visual primitives, including color and texture; hierarchical grouping and learning strategies in the classification process; the ability to deal with rather general objects in uncontrolled configurations and contexts. We illustrate these properties with three case studies: one demonstrating the use of color and texture descriptors; one learning scenery concepts using grouped features; and one demonstrating a possible application domain in detecting naked people in a scene.

This paper demonstrates an automatic system for telling whether there are naked people present in an image. The approach combines color and texture properties to obtain a mask for skin regions, which is shown to be effective for a wide range of shades and colors of skin. These skin regions are then fed to a specialized grouper, which attempts to group a human figure using geometric constraints on human structure. This approach introduces a new view of object recognition, where an object model is an organized collection of grouping hints obtained from a combination of constraints on color and texture and constraints on geometric properties such as the structure of individual parts and the relationships between parts. The system demonstrates excellent performance on a test set of 565 uncontrolled images of naked people, mostly obtained from the internet, and 4289 assorted control images, drawn from a wide collection of sources.

Images taken with wide angle and inexpensive medium angle lenses show substantial distortion, which will cause many computer vision applications to malfunction. Parametric calibration algorithms cannot handle the wide range of possible distortion functions and they require a known image center, expensive equipment, and/or estimation of unneeded parameters. We present a new algorithm to remove distortion. It is similar to the plumb line method but it: (a) uses images of spheres rather than lines, (b) corrects to stereographic rather than perspective projection, (c) calibrates aspect ratio, and (d) uses a nonparametric distortion model.

Presents a method for calibrating intrinsic and extrinsic camera parameters. This algorithm can easily be modified by other users to suit their particular calibration needs, without requiring a high-precision calibration target or complicated linear algebra. The algorithm uses controlled motions and a single light source to simulate calibration targets in convenient 3D locations. These convenient calibration targets enable us to simplify the calibration algorithm and gather dense data for lens distortion. Dense data makes the distortion correction more accurate than traditional low-order polynomial fits, and allows us to calibrate wide-angle lenses (>70/spl deg/ field of view).>

This paper presents a new algorithm for locating the boundaries of textured regions (both step changes and outliers) using a robust estimator. Previous robust image filters perform poorly on binary images, blur edges, round corners, and run slowly. The author avoids artifacts on binary images by modelling them as continuous and interpolating values. Information is combined directly between non-adjacent locations to prevent blurring. Corners are sharpened by relabelling mis-classified pixels. The algorithm is made as fast as a Marr-Hildreth edge finder by restructuring the estimator as a series of 2D image operations, using new multi-ring order statistic operators, and running most of the estimator on a randomly sampled image.>

This work illustrates and explains various artifacts in the output of five finite difference edge finders, those of J.F. Canny (1983, 1986), R.A. Boie et al. (1986) and R.A. Boie and I.J. Cox (1987), and three variations on that of D. Marr and E.C. Hildreth (1980), reimplemented with a common output format and method of noise suppression. These artifacts include gaps in boundaries, spurious boundaries, and deformation of region shape.>

A finite difference edge finder in which the finite difference is computed at a range of widths, i.e. a range of distances between data points, is introduced. Wide operators report low-amplitude responses more reliably than narrow operators, so if wide operators are used to fill gaps in narrow operator responses, each operator can be restricted to report only statistically reliable responses without losing many real features. This sharply reduces the noise in the final output. Theoretical bounds on spurious responses in the finite difference outputs, given only weak assumptions about the signal and noise, are presented. The expected response of the edge finder to an ideal straight step edge is also analyzed. These performance measures are compared with those of a standard algorithm based on Gaussian smoothing and those of a second algorithm that also considers the spatial structure of noise. The algorithms prove equally good at suppressing noise, but are better able to detect faint or blurred features. These predictions are confirmed by empirical tests on real images.>

This paper presents a fast algorithm for computing local symmetry descriptions of region shape. Like previous algorithms, it extracts groups of edge points tangent to a common circle. However, by examining the number of distinct points of tangency, the new algorithm separates groups belonging to round regions from those belonging to elongated regions. Thus, when these relations are connected to form extended regions, round and elongated regions can be processed differently. The new implementation uses an edge-tracking algorithm to build extended regions. This handles the effects of limited precision and shape irregularities better than axistracking methods. In particular, Blum's idea of locating SAT branch points can be converted into a practical method of detecting locations at which three or more regions join. By combining a new density constraint with constraints used previously, the output and much of the processing is made linear in the image area. One method of representing region shapes is to extract symmetry sets, 1 i.e. sets of edge points tangent to a common circle (the symmetry circle), and then group these sets into extended regions (Figure 1). These local symmetry relationships form a intermediate representation between raw edge points and descriptions of whole regions. Thus, a wider variety of regions can be identified than with whole region models as in [1].

Many edge finders extract the signs of finite differences of image intensity values. Camera noise renders many of these signs unreliable. Previous algorithms for reducing noise are difficult to analyze, fail to detect faint or closely packed features, or handle restricted classes of features. The author proposes taking finite differences with a range of separations between data points, and choosing the narrowest response with statistically reliable sign. Fine detail is then detected by narrow operators. Faint features are filled in by wide operators, which can more reliably distinguish low-amplitude boundaries from noise. It is shown, both theoretically and empirically, that this method out-performs traditional Gaussian smoothing. Measurements of noise in a real camera system are also presented.>

The paper describes a new approach to representing space and time for practical reasoning, based on space-filling cells. Unlike Rn, the new models can represent a bounded region of space using only a finite number of cells, so they can be manipulated directly. Unlike Zn, they have useful notions of function continuity and region ‘connectedness’. The topology of space is allowed to depend on the situation being represented, accounting for sharp changes in function values and lack of connectedness across object boundaries. Algorithms based on this model of space are neither purely region based not purely boundary based, but a blend of the two. This new style of algorithm design is illustrated by a new program for finding edges in grey-scale images. Although the program is based on a fairly conventional second directional difference operator, it can detect fine texture in the presence of camera noise, produce connected boundaries around sharp corners and return thin boundaries without ‘feathering’. New algorithms are presented for combining directional differences, suppressing the effects of camera noise, reconstructing image intensities from the second difference values and merging results from different scales (including suppression of spurious boundaries in staircase patterns).

Linguists, including computational linguists, have always been fond of talking about trees. In this paper, we outline a theory of linguistic structure which talks about talking about trees; we call this theory Description theory (D-theory). While important issues must be resolved before a complete picture of D-theory emerges (and also before we can build programs which utilize it), we believe that this theory will ultimately provide a framework for explaining the syntax and semantics of natural language in a manner which is intrinsically computational. This paper will focus primarily on one set of motivations for this theory, those engendered by attempts to handle certain syntactic phenomena within the framework of deterministic parsing.