The Cognitive Basis of Visual Object Perception - Essay Example

The Cognitive Basis of Visual Object Perception INTRODUCTION Visual perception, particularly the recognition of objects is an immediate but complex process. Objects are defined as solid bodies of substance which show “spacio-temporal continuity, cohere within their boundaries when they move, and move only when contacted by other object” (Spelke et al 1995, p.169). According to perception psychologists, object perception is a broad as well as narrow term in the field of vision and cognition. Broadly, the term is applicable to both animate and inanimate objects, and includes: a combination of elements within and a uniting of elements in different groups in the visual output; formation of shape and 3-D structure to some of those elements; facilitation of objects seen earlier to be recognized; and evaluation of the way in which attention is focused on the objects. It includes pictured 2-D and 3-D objects, and illusory objects. Contrastinglly, in the narrow sense the term excludes “many of the conceptual or judgmental processes necessary to distinguish real objects from other entities” (Goldstein 2001, p.170). Thesis Statement: The purpose of this paper is to compare and contrast David Marr’s structural description and computational theories, with Irving Biederman’s recognition-by-components theory of object perception. DISCUSSION The Cognitive Basis of Visual Object Perception Object perception includes both sensory as well as cognitive or mental perception. On the basis of cognitive perception, two people looking at the same object can observe or recognize very different elements about it (Osherson et al 1995). “The psychological study of visual perception grew out of the philosophy of mind” (Bruce et al 2003, p.77) which believes that increased knowledge of the world is acquired through observation and recognition of objects in the environment. Emerging from the artificial intelligence tradition is Marr’s 1982 theoretical approach to visual perception. Advanced research in the domain of visual perception has been based on Marr’s concept. Biederman, 1987 developed the recognition-by-components (RBC) theory of human object perception that is related to early ideas from Marr and other perception psychologists. However, Biederman has revealed evidence of some fundamental differences through pschological experiments conducted by him. Marr’s Structural Description and Computational Theories of Object Perception The first workable structural description theory for human object perception was put forth in 1978 by David Marr and Keith Nishihara. This theory advances that the different sections of an object such as a cat’s leg are “represented by 3-D primitives called generalized cones which specified arbitrary 3-D shapes with a set of parameters” (Goldstein 2009, p.645). For instance, a cylinder can be created by dragging a circular cross-section along a straight line. The circle develops the outline of the cylinder, whose main axis is the straight line. Similarly, a rectangular cross-section when dragged along a straight line, forms the outline of a brick. Marr’s computational theory is related to conventional ideas of object perception, on the basis of which visual perception is the conversion of an arrangement of light on the retina into knowledge of the visual world. Marr concept related to the cognitive and artificial intelligence traditions identifies both image and awareness as representations of the world. The cognitive representation is in neutral grey or spans colours in low intensity and light tints; and in artificial intelligence the depiction is “a symbolic specification of the positions, motions and identities of surrounding objects” (Bruce et al 2003, p.80). This hypothesis has to further find out how the first representation can be refined to obtain the second. Marr used a distinctive method to solve this problem. Biederman’s Recognition by Components (RBC) Theory Subsequent to Marr and Nishihara’s ground-breaking 1978 structural description theory, Biederman proposed in 1987 another leading structural description theory: recognition by components (RBC). Figure 1. Geons are Object Primitives in Biederman’s Study (Irani and Ware 2000, p.3) As seen in Figure 1. above, Geons are the components of objects. They form different objects by bonding in different ways, as illustrated above by geons 3 and 5. Biederman states that objects are cognitively portrayed by a set of 36 components and their association in the dimensions of space. He termed them “geons” for geometrical ions; these were a subset of the generalized cones put forward by Marr and Nishihara. By forming structures on the basis of combinations of geons, commonplace objects such as a cup, pail or table can be created (Goldstein 2009). Biederman’s recognition by components (RBC) theory is developed on Marr and Nishihara’s structural description theory, using two innovative techniques. First, unlike generalized cones, geons differed only qualitatively from each other. For example, a geon’s axis can be either straight or curved, while generalized cones can have any degree of curvature theoretically. Biederman’s second innovation was to put forward a shorter method of disseminating images into geons. According to RBC theory, geons are obtained from non-accidental characteristics which are of edges in an image, such as lines, which are associated with properties of outlines in the world (Goldstein 2009). To understand non-accidental attributes, looking at a box from various angles is an example. Most of the views reveal three sides of the box, with a Y-junction in the corner. A non-accidental property is this two-dimensional junction, which is associated with a three-dimensional corner. From some of the viewpoints, as in a view of one side of the box, the Y-junction is invisible to the viewer. Hence it is uncertain whether such an image is a three-dimensional box or a two-dimensional rectangle. However, these accidental viewpoints occur rarely as compared to non-accidental viewpoints. “Thus, non-accidental properties are highly robust (though not entirely invariant) to changes in viewpoint, viewing distance and illumination” (Goldstein 2009, p.646). Comparison of Marr’s and Biederman’s Theories Marr’s argument was for a segmental and linked theory of visual processing composed of the theory involved in the design of efficient and robust systems in artificial intelligence, and from empirical outcomes in neuropsychology. An example of how one mental faculty can fail, while others continue to function normally is seen in the occurrence of agnosia caused by brain damage, where the patient is unable to identify objects while their other visual abilities remain normal. Thus, according to Marr, visual processing is segmental; the module can be broken down into different subprocesses, each of which changes one portrayal of the object into another. According to Bruce et al (2003), the process of vision is accomplished by a sequential functioning of numerous modules. The first module produces the primal sketch depicting the changes in light intensity taking place on the image. Further, “the primal sketch also organizes these local descriptions of intensity change into a 2-D representation of image regions and the boundaries between them” (Bruce et al 2003, p.80). Subsequently, the layout of visible object surfaces, their distances and positions in relation to the observer are denoted in a 2 ½ D sketch. Finally, using this information, 3-D model representations are created; these highlight the solid shapes of objects and help to identify objects based on their portrayal registered in the visual memory. In the case of Biederman’s RBC theory of object perception, two to three geons are sufficient to represent many objects. Psychologists state that at the basic level, RBC theory accounts for recognizing objects. For instance, the term “snake” is the most primary term for the species, “python” is at a more specific subordinate level, while “reptile” is the general superordinate term (Goldstein 2009). Instead of the structural descriptions proposed by RBC theory, qualitatively different depictions are needed, for the recognition of objects at the subordinate level. For instance, Biederman believes that face recognition is related to “the recognition of fine metric details such as the distance between eyes, rather than geons” (Goldstein 2009, p.646). A second component of Marr’s theory relates to the belief that each processing stage is developed on the representation of the object created at the previous stage, thus ensuring a one-way flow of information starting from a neutral, grey depiction of the object. This opposes the conventional empirist concept in which perception takes place by using previously acquired knowledge of the world to comprehend the retinal image (Bruce et al 2003, p.80). Similar to Marr and Nishihara, according to Biederman, the first stage of object description “involves the segmentation of the occluding contour at regions of sharp concavity” (Bruce et al 2003, p.282). By this method, the various sections of the contour can be related to the depictions of the primitive object shapes, or geons. Then, the features and patterning of the geons can be related to the structural models of objects. The depiction of each known object is a structural version of the parts from which it is developed, including the particulars of size and direction. More than one structural model of an object may be registered in the visual memory, when objects of the same grouping have different forms, as seen in various types of chairs. The main difference between Biederman’s theory and that of Marr and Nishihara is the concept that geons are characterized by features which remain the same from different perspectives. According to this theory, occluding contours need not be used to view a three-dimensional axis-based form. On the other hand, each different kind of geon has its own important features in the 2-D primal sketch level depiction. Thus, unlike Marr’s, in Biederman’s theory object recognition is possible, directly from the 2-D primal sketch, and the building of a 3-D shape depiction is unrequired. Further, new evidence from Biederman’s work indicates the significance of concavities in denoting the partial frameworks of objects. The importance of concavity in object shape perception among chimpanzees was studied by Matsuno and Tomonaga (2007). They found that shape representation in chimpanzees is similar to that of humans’ object perception, in which concave features are important for reconstructing three-dimensional structures from two-dimensional images. Evidence from research conducted by Cooper et al (1992) reveals that object perception by humans on the basis of shape, does not vary with changes of orientation in depth, including parts occlusion, size and changes in the visual field. To solve the invariance problem, Biederman advanced the principle of componential recovery. The component geons are vital for identifying the objects. Since geons are obtained from non-accidental, highly stable properties, the same geon can be obtained from many different viewpoints, distance between object and viewer, and lighting. Conclusive evidence from Biederman’s research on contour deletion confirm the recognition-by-components theory. From the line drawings of commonplace objects such as a mug, he deleted contours that could be used to obtain geons, for example at junctions, “or deleted the same amount of contours from other sections that could not be used to recover geons” (Goldstein 2009, p.646). The evidence indicated that the observers could readily name line drawings in which the junctions were intact, but had difficulty naming line drawings with junctions erased. The above method of Biederman to solve the invariance problem is different from Marr’s attempts to solve the invariance problem by recovering view invariant 3-D models from images. More complex 3-D shapes can be created by dragging different 2-D cross-sections across varying axes. To obtain 3-D generalized cones from 2-D images, the axes of the main parts of an object can be determined from its outline. Generalized cones and their spatial positioning can be deduced using these axes. Perception and recognition would occur by “matching the structural description recovered from the image to those stored in visual memory” (Goldstein 2009, p.645). CONCLUSION This paper has highlighted object perception in relation to vision and cognition in the field of psychology. David Marr’s structural description theory, 1978 and computational approach, 1982 has been compared and contrasted with Irving Biederman’s 1987 recognition-by-components (RBC) theory. It was found that there are similarities as well as differences in the two theories of object perception. According to Marr the ability to recognize objects relates to information being converted into a different representation based on the previous stage, before being transmitted to the next stage. Marr’s computational theory “expresses the specific and very powerful idea that the first stage in understanding perception is to identify the information that a perceiver needs from the world” (Bruce et al 2003, p.81). Biederman’s recognition-of-components (RBC) theory underscores geons as the components of objects, which facilitate object perception. Marr’s and Biederman’s theories are similar in the fact that both state that the first stage of object perception takes into account the concavity of objects. There are differences in the two theories as in Marr’s argument that complete three-dimensional models of objects are recovered, while Biederman states that object representations are three-dimensional but the features may be different depending on the viewpoint. BIBLIOGRAPHY Bruce, V., Green, P.R. and Georgeson, M.A. (2003). Visual perception: Physiology and ecology. Edition 4. Great Britain: Psychology Press. Cooper, E.E., Biederman, I. and Hummel, J.E. (1992). Metric invariance in object recognition: A review and additional evidence. Canadian Journal of Psychology, 46 (2): pp.191-214. Goldstein, E.B. (2009). Encyclopedia of perception. London: Sage Publications. Goldstein, E.B. (2001). Blackwell handbook of perception. Edition 4. London: Wiley- Blackwell. Irani, P. and Ware, C. (2000). Diagrams based on structural object perception. Retrieved on 22nd April, 2010 from: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.26.4171 Matsuno, T. and Tomonaga, M. (2007). An advantage for concavities in shape perception by chimpanzees (Pan troglodytes). Behavioural processes, 75 (3): pp. 253-258. Osherson, D.N., Kosslyn, S.M. and Gleitman, L.R. (1995). An invitation to cognitive science: Visual cognition. Edition 2. Massachusetts: MIT Press. Spelke, E.S., Gutheil, G. and Van de Walle, G. (1995). The development of object perception. In S.M. Kosslyn and D.M. Osherson (Eds.). Visual cognition: An invitation to cognitive science. Edition 2. Massachusetts: MIT Press. Read More