A new method for evaluating edge detection algorithms is presented and applied to measure the relative performance of algorithms by Canny, Nalwa-Binford, Iverson-Zucker, Bergholm, and Rothwell. The basic measure of performance is a visual rating score which indicates the perceived quality of the edges for identifying an object. The process of evaluating edge detection algorithms with this performance measure requires the collection of a set of gray-scale images, optimizing the input parameters for each algorithm, conducting visual evaluation experiments and applying statistical analysis methods. The novel aspect of this work is the use of a visual task and real images of complex scenes in evaluating edge detectors. The method is appealing because, by definition, the results agree with visual evaluations of the edge images. Index Terms-Experimental comparison of algorithms, edge detector comparison, low level processing, performance evaluation, analysis of variance, human rating.

Calibration of comprehension is the correlation between subjective assessments of knowledge gained from reading and performance on an objective test. Contrary to intuition, this correlation is typically close to zero. This article is structured around four points concerning calibration of comprehension. First, poor calibration of comprehension is the rule, rather than the exception, a fact that has been repeatedly demonstrated in our laboratory and in others. Poor calibration is also typical in at least one other domain: solving insight problems. The high levels of calibration reported in studies on the calibration of probabilities and feeling-of-knowing research may be dependent on using feedback from taking the test to assess the probability of correct performance on the test.Second, we present two experiments that demonstrate that poor calibration of comprehension is not associated with a particular type of performance test but is found with inference tests, verbatim recognition tests, and idea recognition tests. For the most part, poor calibration is found when the test is given immediately after reading as well as when the test is given after a delay. Also, we demonstrate that poor calibration cannot be attributed to unreliable testing procedures.Third, the evidence from three experiments indicates that a likely reason for poor calibration is that subjects assess familiarity with the general domain of a text instead of assessing knowledge gained from a particular text. Assessing domain familiarity is probably easier than assessing knowledge gained from a particular text. Also, under some conditions, applying a domain familiarity strategy does result in spurious calibration, which thereby reinforces application of the strategy.Fourth, we demonstrate that calibration of comprehension can be enhanced if subjects are given a pretest that provides (self-generated) feedback. Even this ability is limited, however. Calibration is enhanced only when the processes and knowledge tapped by the pretest are closely related to the processes and knowledge required on the criterion test. Under these conditions, subjects apparently use feedback from the pretest to predict criterion test performance with a modest degree of accuracy. We briefly discuss the implications of these results for theories of representation of knowledge gained from reading.In preparing for a test of learning, a rational strategy is to our claim that beliefs about how much has been learned are study until one believes that the material is learned. Studying often uncorretated with performance on a test of comprehenfor less time is risky; studying for more time may be wasteful.sion. Second, we will demonstrate that the lack of correlation is For this strategy to be effective, however, beliefs and judgments not due to some methodological artifact, but that it is represenabout how much has been learned must be calibrated. That is, tative of a wide range of situations. Third, we present data supthese beliefs must be correlated with performance. Unfortu-porti...

No abstract

Observers responded to full-color images of scenes by indicating which of two critical objects was closer in the pictorial space. These target images were preceded by prime images of the same scene sans critical objects, or by control primes or different-scene primes. Reaction times were faster following same-scene primes than following the various control and different-scene primes. Same-scene facilitation was obtained with color primes, line-drawing primes, and primes with shifted views. The effect occurred with natural scenes having gist and simple artificial scenes having little or no gist. The results indicate that prime-induced representations influence the perception of spatial layout in pictures.

No abstract

A time course contingency is the modification of later phases of object recognition contingent upon stimulus information extracted earlier in processing. It can increase the efficiency of later processing and reduce computational burdens. This idea was instantiated within a global-to-local model and supported in 4 integration priming experiments, in which primes and target objects were presented briefly and then masked. In Experiments 1-3, global and coarse-grained common-feature primes presented early in processing facilitated discriminations between similarly shaped objects, even though they provided no discrimination-relevant information. In Experiment 4, global primes were more effective than local primes early in processing, whereas local primes were more effective than global primes late in processing.An important goal in the study of perceptual identification is to specify the time course of the relevant processes. In this endeavor, marked success has been achieved with pleasingly simple feature accumulation models, in which features are extracted over time and continuously activate form-level representations in a passive, bottom-up manner. For example, such models account for a wide variety of data on pattern and word perception (e.g., Estes, 1978;Keren & Baggen, 1981; Massaro & Sanocki, in press;Oden, 1979;Sanocki, 1990Sanocki, , 1991bShibuya & Bundesen, 1988;Townsend, 1981).The assumption of a bottom-up flow of information can be elaborated on to begin to account for object identification (e.g., Biederman, 1987). However, the adequacy of a bottom-up approach becomes questionable when the computational burdens of identifying three-dimensional objects are considered. Objects vary in size, orientation, and details of instances. They have many features that must be bound together, and objects with moving parts can vary greatly in shape. Any object can appear under different lighting conditions or be partially occluded. There is evidence indicating that the perceptual system can be sensitive to details of instances of letters (e.g., Sanocki, 1987Sanocki, , 1988Sanocki, , 1990Sanocki, , 1991b and objects (e.g., Jacoby, Baker, & Brooks, 1989;Price & Humphreys, 1989) and to differing object orientations (e.g., Jolicoeur, 1985;Palmer, Rosch, & Chase, 1981). This sen- Correspondence concerning this article should be addressed to Thomas Sanocki, Department of Psychology, BEH 339, University of South Florida, Tampa, Florida 33620-8200. Electronic mail may be sent to sanocki@figment.csee.usf.edu. sitivity may extend to differing part configurations and sufficiently different lighting conditions. If during object identification the perceptual system considered such factors for an unconstrained set of alternatives, the enormous number of combinations of stimulus features and feature-object mappings would create combinatorial explosion.These considerations support hypotheses that are not strictly bottom-up in nature. Of interest is the possibility that computational burdens may be reduced by using early stimulus info...

The representation of visual information about letters is proposed to be highly systematic, involving not only abstract information that is invariant across type faces (or fonts), but also a number of parameters whose values are determined by the current font The system exploits regularities that are characteristic of letters and fonts by becoming tuned to the details of the font This should result in efficient letter perception when the stimuli are regular (when all of the letters are of a consistent font), but not when the stimuli are irregular (when the letters are from a variety of fonts) The prediction of faster processing with a regular font, as compared with a mixed font, was examined in three experiments requiring the recognition of four-letter strings Experiment 1 confirmed the prediction, and Experiment 2 replicated the effect with the number of "features" equated across conditions. Experiment 3 showed that the disadvantage for a mixture of fonts is related to how much the representational system must be adjusted to handle the different fonts A central issue in cognitive and perceptual psychology is how familiar objects are represented and perceived. This article concerns the representation and perception of letters, which are representative of more complex objects in that they vary considerably in appearance from instance to instance: The actual form of a letter depends on the type face, or font. Therefore, it is necessary for models of letter perception to specify how the perceptual system maps letters of different fonts onto the appropriate abstract letter codes. The purpose of the present research was twofold, first, to begin developing a new kind of model of letter perception, one that uses the idea of a structural network, and, second, to test this model against a class of simpler models by examining how perceptual representations might become systematically tuned for a particular font.A useful general approach to the recognition problem is to assume that letters are represented and perceived as sets of fea-

The authors argue that the concept of "edges" as used in current research on object recognition obscures the significant difficulties involved in interpreting stimulus information. Edges have sometimes been operationalized as line drawings, which can be an invalid and misleading practice. A new method for evaluating the utility of edge information, operationalized as the outputs of a local, signal-based edge extractor, is introduced. With 1-s exposures, the accuracy of identifying objects in the edge images was found to be less than half that with color photographs. Therefore, edges are far from being sufficient for object recognition. Alternative approaches to the problem of interpreting stimulus information are discussed.