Animation CS 551 / 651 Hodgins et al., 1998 Perception of Physically Simulated Humans Perception How should we render our objects? • To what end? – Verisimilitude – Mechanical accuracy – Impressionism • What about motion? Is simplicity better? Advantages • Abstraction is easier • Obfuscations are removed Disadvantages • Complementary features are removed – Edges, critical features • May look “wrong” Is complexity better? Advantages • Details provide perceptual cues • This is the way we perceive things in real world Disadvantages • Difficult to get the details right • May distract from basic motion We have no idea… We turn to different experts • Psychologists • Automated computer vision Psychologists Kubovy and Proffitt @ UVa • Perception of patterned dot animations – Models of perception • Perception as it relates to action – We perceive because it helps us to act Attacks the perception question within welldefined psychological models Computer Vision Martin and Acton @UVa • Low-level vision – How do we detect edges, shadows, primitives • High-level vision – How do we compose “things” from primitives Still no solid answers Vision and psychology provide models of perception that influence graphics Graphics permits isolated experimentation with perception models The three fields move forward together What’s amazing about us? Perceiving friends • Just two moving lights on ankles is enough • Just two seconds is required – Johansson (1973) Vanrie & Verfaillie (2004) Perceiving pendula • Humans thought moving dots were connected via flexible bar, not rigid pendulum Hodgins’ comparison Is there a difference? Observational tests Torso rotation • Keep head looking forward, but rotate torso and arms Arm Motion • Make arm swing more forward / backward – Adjust dynamics accordingly – How much? Noise • Randomly perturb joint angles (waist, shoulders, neck) – No dynamics – How much? How the simulation works Experimental protocol Watch animations in pairs • 4 seconds of one then 4 seconds of a mate Indicate similarity or difference within pair • forced choice • Could you forget what first looked like? Approx. 25 people per condition Varied the order • Avoids ordering effects (learning during experiment) Experimental protocol Animations rendered in same way • Could this have made a difference? – Is there a rendering that is conducive to stick figures? – What tricks would people use to identify motions? Played from VHS at 30 fps • Can’t have any effects from rendering blips Results On average, people were better with manH Results But how did rendering affect each person’s ability? There’s a trick! Take-away messages Don’t read too much into these results • Each experiment may be different • More detailed model was also more human-like Standardization of animation environments might be good for comparison • Difficult to compare improvements from year to year What else matters? Camera movement Ground plane Motion blur Secondary motion (clothing / hair) Shadows Additional commentary Experiments are essential for graphics • Yet rarely conducted • How is graphics evaluated? – The SIGGRAPH “aaahhhh” factor Additional commentary Creating experiments is dicey business • Have to include psychologists who are experts of experiment design • Make sure enough subjects are included • You need to understand the domain so well that you know the answer before the experiments are complete – Many pretrials were conducted to refine amounts of noise to add (to avoid making it too easy or hard) Follow-up paper Bodenheimer et al., 1999 Eurographics Animation Workshop How does noise influence perception? How to add noise to simulation? Sensors • When the arm reaches angle q, trigger reaction Control gains • How stiff/strong are the muscles Output torques • How regular and well-behaved are the muscles Control parameters • When does the arm swing backwards Output torques Noise inserted here didn’t work well • Instantaneous noise was quickly corrected with subsequent countertorques What kind of noise? Variability of human motion is tied to large movements of the body • Not a random sinusoidal noise function • Not a white noise Experimental scenario Watch 10 movies of varying noise and select the one that looks most “natural”