Effects of lighting on the perception of facial surfaces

Abstract

The problem of variable illumination for object constancy has been largely neglected by "edge-based" theories of object recognition. However, there is evidence that edge-based schemes may not be sufficient for face processing and that shading information may be necessary (Bruce. 1988). Changes in lighting affect the pattern of shading on any three-dimensional object and the aim of this thesis was to investigate the effects of lighting on tasks involving face perception. Effects of lighting are first reported on the perception of the hollow face illusion (Gregory, 1973). The impression of a convex face was found to be stronger when light appeared to be from above, consistent with the importance of shape-from- shading which is thought to incorporate a light-from-above assumption. There was an independent main effect of orientation with the illusion stronger when the face was upright. This confirmed that object knowledge was important in generating the illusion, a conclusion which was confirmed by comparison with a "hollow potato" illusion. There was an effect of light on the inverted face suggesting that the direction of light may generally affect the interpretation of surfaces as convex or concave. It was also argued that there appears to be a general preference for convex interpretations of patterns of shading. The illusion was also found to be stronger when viewed monocularly and this effect was also independent of orientation. This was consistent with the processing of shape information by independent modules with object knowledge acting as a further constraint on the final interpretation. Effects of lighting were next reported on the recognition of shaded representations of facial surfaces, with top lighting facilitating processing. The adverse effects of bottom lighting on the interpretation of facial shape appear to affect within category as well as between category discriminations. Photographic negation was also found to affect recognition performance and it was suggested that its effects may be complimentary to those of bottom lighting in some respects. These effects were reported to be dependent on view. The last set of experiments investigated the effects of lighting and view on a simultaneous face matching task using the same surface representations which required subjects to decide if two images were of the same or different people. Subjects were found to be as much affected by a change in lighting as a change in view, which seems inconsistent with edge-based accounts. Top lighting was also found to facilitate matches across changes in view. When the stimuli were inverted matches across changes in both view and light were poorer, although image differences were the same. In other experiments subjects were found to match better across changes between two directions of top lighting than between directions of bottom lighting, although the extent of the changes were the same, suggesting the importance of top lighting for lighting as well as view invariance. Inverting the stimuli, which also inverts the lighting relative to the observer, disrupted matching across directions of top lighting but facilitated matching between levels of bottom lighting, consistent with the use of shading information. Changes in size were not found to affect matching showing that the effect of lighting was not only because it changes image properties. The effect of lighting was also found to transfer to digitised photographs showing that it was not an artifact of the materials. Lastly effects of lighting were reported when images were presented sequentially showing that the effect was not an artifact of simultaneous presentation. In the final section the effects reported were considered within the framework of theories of object recognition and argued to be inconsistent with invariant features, edge-based or alignment approaches. An alternative scheme employing surface-based primitives derived from shape-from-shuding was developed to account for the pattern of effects and contrasted with an image-based account