Scene Recognition after Active and Passive Learning

Abstract

Most research on visual recognition has been carried out on isolated objects with the main finding being that for certain classes of objects recognition strongly depends on the views learned during training. Recognition of scenes, ie structured environments, is rarely studied, possibly because of the difficulty involved in isolation and control of pertinent cues. We can overcome such problems by using computer graphics to model structured environments where training or learning is facilitated by active explorations with the use of VR technology. We are trying to determine whether there exists the same degree of view-dependence in scenes as has been found for objects. We do this by using a single, sparsely decorated, yet structured room with which subjects familiarise themselves. This learning process can take two forms: either active or passive. In the active case, subjects can manoeuvre in a restricted set of directions in order to find ‘hidden’ coded targets. In the passive case, fifty 2-D views of the room are presented to them in random sequence with some views containing embedded targets which they have to acknowledge. Correct responses and response latencies of eighteen subjects in each condition were recorded in subsequent (old/new) recognition tests. Performance for recognition from familiar directions was similar after active and passive learning (eg approx. 80% hits). However, we found that active learning facilitates recognition from unfamiliar directions ( d' active = 0.96; passive = 0.22). This superior performance after active learning could be due to the increased availability of 3-D information (eg from motion parallax during movement). We are therefore testing this using binocular disparity as a depth cue during passive learning.