Data driven powerpoint templates

Data-Driven

Transcript: Automatic Movement Generation Applications Style Support: Mean values of the p.d.f is a function of factors (style variable) Calculated by multiple regression Supervised Models state duration Allows for defining factors that control the state duration: walking pace What to Evaluate? Data-Driven Human Movement Generation Using Statistical Models High-level factors that affects the movement Gait Identity Effort ... Often referred to ask 'style variable' or 'style space' in the literature Data-Driven Movement Generation Statistical Movement Generation: Overview Style transitions Stride length and speed DBN Learned Style Machine Describing data in statistical terms Each feature vector (frame) contains 60-120 dimensions: multivariate Multivariate probability distribution functions Multivariate Gaussian Parametric Multivariate Gaussian Decomposable Triangulated Graph Mixture Models Data Multiple regression Hidden Semi-Markov Models Wang et al. [2006b] Parameterizing rotations with 3 Degrees of Freedom 3x3 Rotation Matrix Euler Angles 4D Quaternion Exponential Maps Bone Rotations vs Joint Angles Depending on MOCAP format Can be converted to each other Hierarchical representation of joins Rotations relative to parents Temporal: RBMs with direct connections from a number of its past visible units to its hidden and visible units Real-valued visible variables representing joint angles CRBMs as building blocks of Deep Belief Networks Generation based on initialization No style blending or interpolation Factored CRBM Experiments X=strutting+(running-walking) Yamazaki et al. [2005] Semantics Conditional Restricted Boltzmann Machines (CRBM) Yamazaki et al. [2005] Fall October 2013 Applications Discussion Style Extrapolation Pure data modeling No knowledge of human movement involved No physical constraints Having a good data is important! Need a collection of training data Record our own Use a movement database What we can generate mainly depends on what we have as data Multivariate Probability Distribution Functions 3D Rotation of joints or bones Full body movements An skeleton can be used for visualization purposes Joint positions Sources Motion Capture Kinect, Leap Motion Accelerometer (not good for generation) ... Not violating body constraints Aesthetics Realistic/Believable Variance from the training data Is it generation new examples? How different are examples from each other? Prediction How well can the model predict the movements given an initial set of frames Movement Data Physics based Data-driven Machine Learning Models - Differences Style-Content Separation: Motivation Wang et al. [2007] Style-Content Separation: Techniques Movement database with proper annotations Represent a range of different movement semantics Approaches for extracting and representing semantics Choice of the contents and semantics affects the data collection requirements Data source Marker placements for mocap Framerate SHMM http://www.uoguelph.ca/~gwtaylor/thesis/4/ Multifactor Gaussian Process Model Collection of motion capture files Used for training Need to be labeled/annotated for supervised methods Ideally includes Multiple takes of the same movements with same and different subjects Variations of the same movement type: stylized walks Follows a set of proper semantics Wang et al. [2006b] Style Content Separation Data Representation Defining the style dimensions Supervised Unsupervised Incorporating style into the model Conditional Probability Distribution Model Architecture Multivariate Gaussian Correlation between all dimensions Parametric Multivariate Gaussian The mean value of the distribution is affected by a parameter (style variable/factor) Decomposable Triangulated Graph HMM/Mix-SDTG HMM/Mix-SDTG Different models, different architectures, different representational capabilities Hidden Markov Models: Parametric HMM, HSMM Artificial Neural Networks: Feed-forward, CRBM, Factored CRBM Gaussian Process Models ... Specialized learning algorithms Generation Algorithms Probability Distribution Functions Taylor et al. [2006] Wang et al. [2007] Between 60 to 120 Includes root position and orientation 1, 2, or 3 DOFs for each joint/bone Depends on Type of the rotation parametrization Number of joints/bones defined in skeleton Number of markers used (marker based mocap) Dimensionality Reduction Algorithms Principal Component Analysis Hard to evaluate the works No methodological evaluation method Usually no videos available Not replicable Lack of a comprehensive database Repeated variations of the same movements Valid annotations No semantics defined in current databases or models No study on agent model integration Mostly using motion capture Approaches: Concatenation Blending and interpolation Transforming existing movements Statistical models Style Machines Multifactor Gaussian Process Model: Experiments Extending a MOCAP database Can be used by animators Useful for other animation technique Cost and time Direct Animation Current

Data Driven Powerpoint Templates

Create your presentation by reusing a template from our community or transition your PowerPoint deck into a visually compelling Prezi presentation.

DATA DRIVEN

Data Driven

Data Driven

data driven

Data-Driven

Data-Driven

Data Driven

Data driven

Explore our templates for more presentation inspiration

Customer Stories - Office

Sales KickOff - Rocket

Quarterly Business Review - Downtown

Marketing Proposal Template for Powerful Presentations | Prezi