Abstract
Background: Faces are important stimuli in social interactions, but the perception of bodies may also play an important role in person identification and inference of emotional state. Adaptation has proven a useful means of exploring face representations in the visual system, and can inform us of the nature of body representations. Body aftereffects may be particularly useful for studying invariance in object representations, as they can be subjected to more drastic manipulations of pose. Objective: Our goals were to determine if body aftereffects could be obtained, and if so, to what degree these show viewpoint and pose invariance. Methods: Headless body images were generated from a realistic 3-D mesh model of the human body created from laser range scans of over 2000 people. Statistical machine learning methods were used to factor body shape variations due to identity from those due to pose. By varying the parameters of the model we can generate realistic body shapes in any pose and viewpoint. In experiment 1, we used different viewpoints of an upright body for adapting images and frontal views for test stimuli. In experiment 2, we used the same frontal view of upright bodies as test stimuli, but compared adaptation with the same upright pose to that with adapting body stimuli in different poses. Results: We found aftereffects for upright bodies that remained significant across viewpoint changes. In contrast, there was minimal transfer of adaptation across changes in pose. Conclusion: Body aftereffects show significant transfer across viewpoint, in contrast to the sharp decreases in face adaptation with change in viewpoint that have been previously reported. Lack of transfer across pose indicates a significant limitation to the invariance of body representations, however.
NSERC Discovery Grant RGPIN 355879-08.