Selection of Best Reference Objects in Object Localizations

The quality of an object localization depends essentially on the adequate selection of a suitable reference. In most computational approaches developed so far only the distance between the located object and a potential reference object has been used as a decision criterion. However many other criteria have to be considered for a cognitively plausible selection of adequate reference points. In this paper we investigate how object and context dependent properties, like referentiality, visual salience, functional dependencies, or prior knowledge, influence the quality of a reference object. Each factor is quantitatively determined and scaled by relevance to a certain context. The scaling permits the necessary comparability of the different quality criteria. Finally, on the basis of these factors a computational model is presented which permits a context dependent determination of the best reference object in a particular spatial configuration. Introduction "Where is object A positioned~. ". The answer to spatial queries like this requires locating a known object A with reference to another (reference) object. The localization of a particular object is an often required procedure in many applications dealing with the domain space. The applications range from general spatial information systems, like geographic information systems or driver information systems, and systems using multi-modal instructions, to applications in Virtual Reality. Understanding the use of spatial references permits the development of better interfaces between systems and human users in the spatial domain. Most algorithms developed so far in this context use very simple mechanisms for the determination of adequate reference objects (e.g., (Carsten & Janson 1985; Wazinski 1992). The distance from the located object (LO) to the reference object (RO) is often the only terion used. However many other crucial factors which depend either directly on the objects or on context factors need also to be considered. The most easily recalled attributes of a region are typically referred to as landmarks. They are used to denote distinguishing features of a route or a region (Lynch 1960; Appleyard 1969; Downs & Stea 1973; Siegel & White 1975). In the first case they support navigational decisions, whereas in the second case landmarks allow for the maintenance of general geographical orientation. Recognizability (Lynch 1960), use (Appleyard 1969), and cultural meaning (Moore 1979) have been emphasized as being the key factors for the relative landmark status of a place. The metaphor cognitive map is usually used to describe mental representations of environments, including landmarks, knowledge about the spatial relations between them, routes and metric survey information (G~irling & Golledge 1989). However the cognitive map seems to be systematically distorted and potentially contradictory (see e.g., (Stevens & Coupe 1978; Sadalla, Burroughs, & Staplin 1980; Tversky 1981; Holyoak & Mah 1982; Hirtle & Jonides 1985)), and thus not easily reconcilable in a map-like structure (Tversky 1993). This led Tversky to propose the term cognitive collage as often being a more appropriate metaphor for environmental knowledge than cognitive map. Some landmarks function as spatial reference points, points that serve as the basis for the spatial location of other non-reference points. The concept of spatial reference points implies that the position of a large set of non-reference locations in a particular region is defined in terms of the position of a smaller set of reference locations (Sadalla, Burroughs, & Staplin 1980). Places, known as reference points, are relatively better known and serve to define the location of adjacent points. According to the findings of Rosch (1975) about the existence of asymmetries in similarity judgments between semantic reference points and non-reference points, Sadalla, Burroughs, and Staplin were able to measure asymmetries of cognitive distances between spatial reference points and non-reference points. For example, subjects were able to indicate the proximity of reference points faster than they could indicate the proximity of equidistant non-reference points, and that the direction of particular target locations is more quickly verified relative to reference locations, than relative to non-reference locations. An experimental cluster 23 From: AAAI Technical Report SS-96-03. Compilation copyright © 1996, AAAI (www.aaai.org). All rights reserved.