Precomputing Avatar Behavior From Human Motion Data Jehee Lee and Kang Hoon Lee Symposium on Computer Animation ’ 04 Date: 12/5/2006 Reporter: 彭任右

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data2 Outline IntroductionIntroduction State-Action ModelState-Action Model PrecomputingPrecomputing SynthesisSynthesis ExperimentsExperiments DiscussionDiscussion

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data3 Goal Pre-computationPre-computation Interactive ControlInteractive Control Motion DataMotion Data Minimum Runtime CostMinimum Runtime Cost

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data4 Introduction PreprocessingPreprocessing Motion capture dataMotion capture data Lookup tableLookup table InputInput Avatar StateAvatar State Target StateTarget State OutputOutput A sequence of action to the goalA sequence of action to the goal

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data5 State-Action Model

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data6 State-Action Model Avatar StateAvatar State S ActionAction A Target StateTarget State E State-action PairState-action Pair {(S,E), A}

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data7 State Space

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data8 Precomputing Pre-compute which action to take at any situationPre-compute which action to take at any situation Reinforcement learningReinforcement learning Trial-and-errorTrial-and-error Discover which actions tend to increase the long-run sum of rewards in future trialsDiscover which actions tend to increase the long-run sum of rewards in future trials

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data9 Reward Specific to a behaviorSpecific to a behavior Approach the targetApproach the target Throw punches at the targetThrow punches at the target Dynamic ProgrammingDynamic Programming Randomly select one state pairRandomly select one state pair Select the action that gains the highest reward in one stepSelect the action that gains the highest reward in one step Value iterationValue iteration

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data10 Synthesis OptimalOptimal Select the action with the highest valueSelect the action with the highest value MultipleMultiple Select an action with maximum weighted sumSelect an action with maximum weighted sum RandomRandom Identify a small number of preferable actionIdentify a small number of preferable action Select one randomlySelect one randomly

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data11 Experiments Pentium 4 2.4GHzPentium 4 2.4GHz 1GB memory1GB memory Vicon 120 FPS -> 15 FPSVicon 120 FPS -> 15 FPS 8 minutes shadowboxing8 minutes shadowboxing 20 different combination of punches20 different combination of punches 5 * 5 capture region5 * 5 capture region

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data12 Data Annotation ContactContact If ankle or toe is close to the ground and its velocity is below some thresholdIf ankle or toe is close to the ground and its velocity is below some threshold Effective HittingEffective Hitting The magnitude of the fist ’ s velocity projected onto the forearm axis is above some thresholdThe magnitude of the fist ’ s velocity projected onto the forearm axis is above some threshold 788 effective hitting points788 effective hitting points

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data13 Graph Construction TransitionTransition If poses at frame i and frame j-1 are similarIf poses at frame i and frame j-1 are similar Both left foot are about to leave the groundBoth left foot are about to leave the ground Strongly Connected ComponentStrongly Connected Component 437 states437 states transitions27072 transitions

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data14 Performance PreprocessingPreprocessing ApproachApproach 1 hour1 hour Throw PunchesThrow Punches 7 hours7 hours Run TimeRun Time 30 animated boxers sparring30 animated boxers sparring 9 second to create 1000 frames with video and sound disabled9 second to create 1000 frames with video and sound disabled

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data15 Results

Alvin\CAIG Lab\CS\NCTUPrecomputing Avatar Behavior From Human Motion Data16 Discussion Find the optimal behaviorFind the optimal behavior Better than on-line search algorithmBetter than on-line search algorithm Difficult to coordinate multiple goalsDifficult to coordinate multiple goals Memory intensiveMemory intensive Apply to other human motionApply to other human motion Collision detection and responseCollision detection and response