Chapter Four: Feature Line Textures
Total Page:16
File Type:pdf, Size:1020Kb
Chapter Four: Feature Line Textures Preceding chapters described the motivation for adding a sparse, opaque texture to an overlaid transparent surface, to help better communicate its shape and relative depth distance from underlying opaque structures while preserving its overall transparent character. The goal of this chapter is to describe the perceptual motivation for selectively opacifying valley and ridge regions and to present an implementation that I developed, independently and concurrently with similar efforts elsewhere, to do this. Inspired by the ability of gifted artists to define a figure with just a few strokes, I would like to define a technique for illustrating layered transparent surfaces, in three dimensions, so that they can be both clearly seen and easily seen through at the same time. My aim is to efficiently and effectively communicate the essential features of the superimposed surface in an intuitively meaningful way, using a clear and simple representation that is appropriate for dynamic viewing conditions, minimizes extraneous detail and allows a largely unobstructed and undistorted view of underlying objects. Research in pictorial representation and image understanding indicates that line drawings, and outline in particular, are a natural and universally understood means for communicating information about objects. A wide range of experiments (cited in [Kennedy 1974]) with animals, children and individuals from a variety of culturally-diverse populations provide evidence that the ability to recognize objects from line drawings is inborn (as opposed to a learned technique). The ease with which we interpret line drawings may suggest an intrinsic relationship between this type of representation and the way our visual system processes and stores visual information. David Marr’s “primal sketch” theory of visual information processing [Marr 1976], for example, is founded on the observation that many of the essential elements of image understanding can be derived from the type of information that is encoded in a line drawing representation of a scene; he postulates that the extraction of this “primal sketch” is a first step in visual information processing, at least for two-dimensional intensity images. Experiments by Biederman and Ju [1988] showed that familiar objects could be identified slightly more quickly and accurately in simple line drawings than in color slides when the images were presented, with masking, for extremely brief (50ms) exposure durations. (The accuracy and speed of object identification was equivalent for the two presentation methods when exposure durations were longer, and the objects could be correctly identified in nearly all of the images when viewing times were extended to 100ms.) These results provide further evidence both for the importance of the information that a line representation can carry and also for the merits of a display mode in which important information is highlighted while less essential detail is removed to improve clarity. Figure 4.1 illustrates some of the sample stimuli used in these object recognition experiments. Although the essential elements of many objects can be completely captured in a line drawing, it’s important to recognize that simple lines (particularly contours or outlines) don’t always provide sufficient information for object representation in general. Biederman and Ju [1988] noted that objects whose projections were structurally very similar (for example a peach and a plum) could not be easily differentiated on the basis of outline or contour information alone. They also observed that certain types of objects — such as hairbrushes — were poorly characterized by their outline. 77 Figure 4.1: Examples of sample stimuli used by Biederman and Ju [1988] in experiments showing slightly lower error rates and faster reaction times for naming (or verifying the identity of) objects represented by line drawings as opposed to color photographs, for very brief, masked exposure durations. Perceptual studies in facial recognition have consistently shown that observers perform poorly on tasks measuring the ability to identify familiar people from line drawings (and particularly poorly when only “outlines” of depth and intensity discontinuity are provided), even though the same individuals can be easily identified in the photographs upon which the line drawings are based. In experiments by Davies et al. [1978], individuals were correctly identified 90% of the time in photographs but only 23% of the time in outline (depth discontinuity and feature boundary) representations and 47% of the time in more detailed line drawings (in which intensity discontinuities were represented in addition to the outline). Subsequent experiments by Bruce et al. [1992] showed that inclusion of even the most basic form of bi-level shading (see figure 4.21-right for an example) could dramatically improve performance on facial recognition Figure 4.2: Examples of photographic vs. line drawing facial recognition task stimuli, generated by [Rhodes et al. 1987]. Left: a black & white photograph. Center: a line tracing of the photograph. Right: a computer-generated caricature derived from the line drawing. 78 tasks. The images in figure 4.2, from [Rhodes et al. 1987], are indicative of the type of sample stimuli used in these kinds of facial recognition experiments. Interestingly, the studies by Rhodes et al. [1987] showed that familiar individuals could be more easily recognized in line drawings when their most atypical facial features were selectively exaggerated, in the style of a caricature. It appears that simple outline drawings — in which lines are used only to mark intensity and depth discontinuities — lack certain essential information required for some higher-level object identification tasks. Although we can often easily categorize an object based on a simple line representation, experiments such as those by Price and Humphreys [1989] indicate that we do use other surface information, particularly shape, texture, or characteristic color information, to make finer distinctions between similar members of the same class. Price and Humphreys [1989] suggest that edge-based and surface-based recognition processes might operate in a cascaded manner, with the effects of the latter being most significant when the information encoded in the former is inadequate for the task. Despite the insufficiency of line drawings for some higher-level object differentiation tasks, it remains evident that line representations can convey a great deal of information about the objects in a scene. For a number of reasons, which will be discussed in greater detail in chapter five, texturing methods in which the surface opacity is explicitly varied according to the surface shading parameters do not appear particular promising; preliminary experiments — described in the appendix — appear to confirm this view. However, a sparse, opaque texture that approximates a “three-dimensional line drawing” might be useful for emphasizing the important features of a transparent surface without unduly occluding underlying objects. Recognizing the roles of various types of lines in visual perception, the question then becomes: how can we best define a set of descriptive lines to clearly and efficiently communicate the essential shape features of a transparent surface in a perceptually intuitive way? 4.1: Silhouette and contour curves Silhouette and contour curves are the two-dimensional projection of points on a surface in 3-space whose surface normal is orthogonal to the line of sight. Silhouette curves form a closed outline around the projected form, while contour curves may appear within the projected form and may be discontinuous. Figure 4.3 illustrates the silhouette and contour lines in a projected image of a simple object, from [Koenderink 1990]. Figure 4.3: Examples of silhouette and contour curves in the projection of a simple object. Left: detailed image of a banana-shaped surface, from [Koenderink 1990]. Center: the silhouette curve. Right: the contour curve. 79 Silhouette and contour curves are ubiquitous in two-dimensional line art and illustration; it’s difficult to imagine a successful line drawing that isn’t based in large part upon these curves. Silhouette and contour lines possess a number of features that make them useful for image understanding and object recognition. Of particular importance, contour lines mark the depth discontinuities in a two-dimensional image, and silhouette lines separate the figure from the ground. One popular method for automatically generating line drawings of three-dimensional models, proposed by Saito and Takahashi [1990], operates directly from this observation, defining the contour lines in a two-dimensional projection by using a gradient operator to locate discontinuities in the depths of the surfaces that map to adjacent pixels in the projected image. Figure 4.4, from [Saito and Takahashi 1990], shows examples of some “perceptually-enhanced” images produced by this method. Figure 4.4: Enhanced images, generated by Saito and Takahashi [1990], in which various types of depth discontinuities, found using an edge operator on a two-dimensional depth map, are highlighted with black or white lines. Left: Upper left quadrant: the original “nut” image. Lower left quadrant: black lines represent the detected locations of zero-order depth discontinuities; white lines represent the detected locations of first-order (slope) discontinuities. Upper right quadrant: lines of zero-