The Effect Of Spatial Properties On View Combination Mechanisms In Scene Recognition

To interact with a dynamic environment, humans alike must be capable of accurately and quickly recognizing objects or scenes despite changes in view that occur through movement of the object or of the observer. Research on the apparent ability of human vision has been largely driven by two classes of cognitive models—Structural description models and View-based models. One of the theories in View-based model elaborating the object recognition from a computational perspective known as view combination suggests that generalization to novel views of an object is accomplished by combining multiple object views represented in memory. That is, the extent to which a novel view of an object can be readily recognized depends on its degree of structural similarity to a set of multiple stored views. Although there have been many empirical studies on the view combination mechanisms of objects recognition in the past leading us to know some factors such as similarity and temporal-spatial continuity between two studied view of object could affect the view combination processing, only a few have investigated how views of scenes are combined across different viewpoints and how the system uses this information in scene recognition.In the present study, based on the improved Friedman (2008)'s paradigm for view combination, we created a virtual-reality setting much like Diwardker's real table layout to explore the view combination mechanism. Considering there are various features intermingled in our surrounding environment, we made a further discrimination in the effect of location information and identity information on view combination in scene recognition.Any of the 3 experiments we reported here consisted of two phases:training phase and testing phase. Participants were allowed to complete a scene recognition task regardless of which perspective the scene was taken from. The virtual table layout was presented from two ground-level perspectives 75°apart in the training phase(e.g.0°training view and 75°training view), and three other perspectives in the testing phase(e.g. interpolated view-one that was between the two training perspectives i.e.37.5°view, extrapolated view-one that was outside of the training range i.e. 112.5°view, and far view-one that was far from the training views i.e.150°view. Both the interpolated view and the extrapolated view were equidistant from one of the training views,) Participants received feedback message in training phase but none in testing phase.In expl,the independent variable of interest was testing view type(training view, interpolated view, extrapolated view and far view) and distractor type(move, switch).The result of experiment 1 showed us subjects recognized interpolated views as fast as the trained views. Both interpolated view and training view were recognized faster than extrapolated view, although extrapolated view and interpolated view were both equidistant from the training view. The same pattern was replicated in error rate in the move condition rather than in the switch condition. The reason for the difference was elaborated intensively in the discussion.In exp2,in order to explore the effect of different spatial properties on the view combination mechanisms, we created two sorts of scenes composed of different spatial information. One sort of two layouts consisted of identical cups arranged irregularly on the virtual desktop, and the corresponding distractor of which is the layout with one object moved to other position. The other sort of two layouts consisted of different cups (of different color, texture, or shape), and the corresponding distractor of which is the layout with two objects switched. The independent variable of interest is also testing view type(training view, interpolated view, extrapolated view and far view) and spatial properties information contained in the scene (location information, identity information). According to the result of Exp2, subjects recognized interpolated views more slowly than the trained views, but faster than extrapolated views in location information condition. In the identity information condition, subjects recognized interpolated views as well as the trained views, but faster than extrapolated views. The same pattern was replicated in error rate in either condition. Until now, we could make the conclusion that view combination effect occurred in either condition. Moreover, the view combination effect in identity information condition is stronger than that in location information condition.In exp3, we brought in the dynamic cues. Other things being equal, we changed the presentation of the layout to examine how subjects use the dynamic cues in experiment. The layout was presented from one perspective for 3 seconds firstly, and then the layout was covered by a black cloth, and in the same time, the virtual camera rotated around the center of the layout for 1s in angular speed of 75°/s. When the camera achieved in the second viewpoint, the layout was presented for 3s once again with the black cloth removed. As a result, in either condition, subjects performed the same way as in experiment 2.Moreover, in identity information condition, subjects recognized interpolated views as well as the trained views, far faster than extrapolated views which seemed that subjects performed much better than did in experiment 2. The result of experiment 3 showed that motion cues did facilitate the view combination effect in identity information condition; in contrast, it seemed that motion cues did not have any facilitation effect in location information condition.In conclusion, this finding of our experiments provided more evidence for view combination in scene recognition, and shed light on how location or identity information influences the view combination process in scene recognition individually. Also, our result brought us more consideration about the contribution of dynamic cues to the effects of spatial properties on view combination.