Transcript motion.ppt
Data Driven Models of Motion Walking the Fine Line Between Performance, Realism and Style Chris White G-Lunch 2007 A Motivating Thought Experiment • You are the lead developer on an upcoming video game. The designer gives you the following requirements: – This game must have tons of fire – Nobody will buy the game if the fire does not look real What do you do? Uncanny Valley Roadmap • • • • Video Textures Motion Graphs Synthesizing Motion from Annotations Group Motion Graphs Video Textures Markov Property of Video Frame l Frame i P(i,j) P(i,k) Frame j Frame m Frame k P(k,m) Probability of a Transition • Frames i, j Is Video really Markovian? Alternative Question: Why is C0 continuity not enough? Dead End Transitions Interesting Open Problems • Relighting? Motion Graphs • Apply Video Textures idea to motion capture data Motion Graph Clip l Clip i P(i,j) P(i,k) Clip j Clip m Clip k P(k,m) Detecting Candidate Transitions • Weighted sum of L2 distance of point clouds of joint positions Candidate Transitions Continued • Find local minima of optimization Bad Transitions • Sinks and Dead Ends • Compute Strongly Connected Components Searching for motion • Random graph walk is only useful as an elaborate screen saver • User supplies function g(w,e) describing additional error of adding edge e to path w • Total error of path is: Downside • Designers/artists do not like specifying their desires as mathematical functions Motion Synthesis from Annotations • Allow animator to paint semantically meaningful timeline for motion Annotating the Motion Database • Generate meaningful vocabulary – For football: Run, Walk, Wave, Jump, Turn Left, Turn Right, Catch, Reach, Carry, Crouch, Stand, Pick up • For each term, designer annotates small subset of database • Support Vector Machine classifies remaining motions Generating Motion • Minimize the objective function: • D(i,A(f)) determines if annotation is met at that frame • C measures the continuity of putting two frames together Optimize using Dynamic Programming Interesting Extensions • Explore the space of motions – We looked at the space of BRDFs last semester