AbstractLittle is presently known about action observation, an important perceptual component of high-level vision. To investigate this aspect of perception, we introduce a two-alternative forced-choice task for observed manipulative actions while varying duration or signal strength by noise injection. We show that accuracy and reaction time in this task can be modeled by a diffusion process for different pairs of action exemplars. Furthermore, discrimination of observed actions is largely viewpoint-independent, cannot be reduced to judgments about the basic components of action: shape and local motion, and requires a minimum duration of about 150–200 ms. These results confirm that action observation is a distinct high-level aspect of visual perception based on temporal integration of visual input generated by moving body parts. This temporal integration distinguishes it from object or scene perception, which require only very brief presentations and are viewpoint-dependent. The applicability of a diffusion model suggests that these aspects of high-level vision differ mainly at the level of the sensory neurons feeding the decision processes.