Abstract
Agents are not omnipotent. Instead, their motions are driven by both their internal intentions and outside constraints (such as an impulsive dog resisting its leash). In many cases, agents are even pulled away from their goals. How does vision interpret such intention-motion dissociations? One answer is that it simply can't, as suggested by studies showing that vision has little tolerance for deviations from perfect goal-directed motions (e.g. Gao et al., 2009). Alternatively, vision can detect intentions from deviated motions, provided it can explain away deviations as causal constraints imposed by the environment. We tested this hypothesis with the Search-For-Chasing task, in which a "wolf" pursues a "sheep" among distractors. The wolf's sheep-directed motion is compromised either by Causal or Non-Causal deviations. Causal deviations are generated by introducing the classic "motion hierarchy" phenomenon (Johansson, 1950) into a chasing display. The hierarchy puts the wolf on a virtual "leash", dragging the wolf away from the sheep-direction by composing the wolf's pursuit with the motion of its superior in the hierarchy. Non-Causal deviations are created by keeping the magnitude of the deviations while destroying the motion hierarchy. In Expt.1, both the wolf and sheep are constrained by a single superior, producing an averaged 45° deviation in the wolf's pursuit. Chasing detection is 21% higher with Causal deviations, compared to Non-Causal deviations. In Expt.2, the wolf is the subordinate of a distractor while the sheep flees freely, producing an averaged 65° deviation. The performance is again 20% higher with Causal-deviations. These results demonstrate that perceived animacy is more robust and flexible than previously suggested, provided the display can be explained by a causal hierarchical structure. We summarize it as a "leash resistance" effect, in which vision intelligently interprets intention-motion dissociations by jointly inferring the agent's internal intentions and outside causal constraints.
Meeting abstract presented at VSS 2018