This theory has motivated the design of a number of experimental studies and it is now relatively well established that our visual system exhibit greater sensitivity to differences in NAPs compared to MPs (see Biederman, 2007, for review).īehaviorally, it has been shown that participants can more accurately distinguish between two objects that differ along an NAP vs. These geons can be differentiated on the basis of differences in NAPs, and generic object categories can be represented as compositions of geons. Briefly, this structural-description theory states that the visual system may encode a finite visual vocabulary of basic 3D shapes called geons. Indeed, NAPs have been the focus of a prominent psychological theory of object recognition called the Recognition-by-Components (RBC) theory ( Biederman, 1987). The stability of NAPs over viewpoints makes them useful for achieving object constancy. For instance, the probability of a curved edge to appear straight because of projection is extremely small and would happen as an “accident” of viewpoint ( Richards et al., 1996). There is a long history of studies related to NAPs in computational vision (see Lowe, 1984, for review): From a theoretical point of view, a visual system needs to focus on the detection of image structures that are unlikely to have arisen by accident. These qualitative properties are known as non-accidental properties (NAPs) and need to be contrasted with their quantitative counterparts known as metric properties (MPs). Conversely, there also exist qualitative shape properties that remain stable across changes in viewpoint, e.g., whether an edge is straight or curved, whether a surface is convex or concave, or whether a cross section ends at a point vs. Properties such as the degree of curvature of an object's contours, its length, or the amount of expansion of a cross section are examples of properties that will be affected by changes in viewpoint. In particular, one may distinguish between those object properties that will remain stable across changes in viewpoint and those that will not (see Figure 1, for an illustration). Object constancy requires the development of visual representations that remain stable across object transformations ( Földiák, 1998). Yet, despite these large intra-class variations, primates are capable of robustly and effortlessly recognizing objects ( Thorpe et al., 1996), vastly outperforming the best existing computer vision systems. Our visual system has to deal with large intra-class variations owing to the effect of 2D and 3D transformations (including translation, scaling and rotation) because small changes in an object's 3D view may yield large changes on its 2D projection on our retinas. Invariant object recognition is a notoriously challenging computational problem ( Marr, 1982). Together, these results suggest that greater NAP sensitivity may arise from experiencing different in-depth rotations of objects. We show that unsupervised learning yields shape-tuning in higher stages with greater sensitivity to differences in NAPs vs. We collected a database of video sequences with objects slowly rotating in-depth in an attempt to mimic sequences viewed during object manipulation by young children during early developmental stages. Here, we test the hypothesis that unsupervised learning of invariances to object transformations may increase the sensitivity to differences in NAPs vs. These models typically assume that object processing is based on populations of view-tuned neurons with distributed symmetrical bell-shaped tuning that are modulated at least as much by differences in MPs as in NAPs. However, previous work has shown that such sensitivity is lacking in multiple-views models of object recognition such as H MAX. Behavioral and electrophysiological studies of shape processing have demonstrated greater sensitivity to differences in NAPs than in MPs. curved contours) and are distinguished from metric properties (MPs) that can change continuously with in-depth object rotation (e.g., aspect ratio, degree of curvature, etc.). Non-accidental properties (NAPs) correspond to image properties that are invariant to changes in viewpoint (e.g., straight vs. 2Brown Institute for Brain Sciences, Providence, RI, USA.1Department of Cognitive, Linguistic, and Psychological Sciences, Brown University, Providence, RI, USA.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |