Spring 2019 CSCI 621: Digital Geometry Processing 13.1 Dynamic Geometry Processing I Hao Li http://cs621.hao-li.com 1
Problem Classification 2
Correspondence Classification 3
Correspondence Classification 4
3-D Reconstruction acquisition registration initial merging alignment data provided by Paramount Pictures and Aguru Images 5
Non-Rigid Registration data provided by Paramount Pictures and Aguru Images 6
Full Reconstruction data provided by Paramount Pictures and Aguru Images 7
Correspondence Classification 8
Dynamic Input Data continuous motion / general deformation 9
Dynamic Input Data Data provided with T. Weise and L. Van Gool 10
Dynamic Input Data momentary motion / articulated deformation 11
Animation Reconstruction 12
Animation Reconstruction 13
Dynamic Shape Reconstruction SIGGRAPH 2009
Template-Based Reconstruction transfer analyze transfer analyze deformation deformation deformation deformation SIGGRAPH 2009
Correspondence Classification 16
Pairwise Correspondence shape & pose / general deformation 17
Statistical Shape Spaces 18
Statistical Shape Spaces 19
Scan Data - Challenges 20
Challenges 21
Correspondence Problem SIGGRAPH 2009
Non-Rigid Registration 23
Pair of 3D Scans target source SIGGRAPH 2009
Correspondences are Lost ? SIGGRAPH 2009
Overlapping Regions are Lost missing data overlapping regions SIGGRAPH 2009
Overlapping Regions are Lost SIGGRAPH 2009
Non-Rigid Registration SIGGRAPH 2009
The Recipe source registration target detect correspond deform overlap
The Challenge ambiguity deformation detect correspond deform overlap
The Challenge ? detect correspond deform overlap
The Challenge detect correspond deform overlap
Observation correspond helps helps detect deform overlap global optimization via local refinement
Iterative Global Optimization correspond detect deform overlap
Iterative Global Optimization closest point correspond detect pruning Robust overlap Non-Rigid ICP global deform optimization no converges? yes relax stiffness
Deformation Model E rigid E smooth de-coupled complexity detail preservation global consistency SIGGRAPH 2009
Non-Linear Energy Minimization E rigid E smooth c i v i [Chen & Medioni ’92] E point E tot E plane E plane E rigid + α point + α rigid + α smooth = non-linear least squares Jacobian is minimization sparse sparse Cholesky factorization Gauss-Newton method that’s it! SIGGRAPH 2009
Summary two scans Sampling Correspondence Correspondence must be robust w.r.t. underlying deformation Weighting Deformation In general: Non-linear problem non-rigid E tot = α fit E fit + α reg E reg registration 38
Summary α rigid → 0 α smooth → 0 Relax Regularization Correspondence Weighting Deformation non-rigid registration • Example with Embedded Deformation Model 39
Symmetries feature matching sampling clustering region matching Source: [Chang and Zwicker 08] 40
Isometry Preservation input data sampling correspondence registration clustering Source: [Huang et al. 08] 41
Dynamic Shape Reconstruction 42
Multi-Frame Reconstruction transfer deformation SIGGRAPH 2009
Geometry and Motion Reconstruction data provided by Stanford and MPI Saarbrücken 44
input data template fitting data provided by Stanford and MPI Saarbrücken 45
More Results Input Scans Reconstruction Textured Reconstruction SIGGRAPH 2009
More Results Input Scans Reconstruction Textured Reconstruction SIGGRAPH 2009
More Results Input Scans Reconstruction Overlaid Scans SIGGRAPH 2009
Template Fitting 49
Initial Alignment template first scan collaboration with MPI Tübingen/Brown University
In Practice: Need Some Correspondences Collaboration with MPI Tübingen
Improving SCAPE non-rigid alignment pose estimation regression sparse/partial matching SCAPE model accurate model Collaboration with MPI Tübingen
Regression Results > 50% more accuracy
Alignment Comparison Collaboration with MPI Tübingen
Alignment Comparison Collaboration with MPI Tübingen
Template Free-Reconstruction 56
Temporally-Coherent [Li et al. ’11] Shape Completion partial data reconstruction partial data reconstruction
Free-Viewpoint Video [Li et al. ’11]
3D Reconstruction 59
Multi-View Capture multi-view stereo multi-view photometric stereo
Single-View Capture [Rusinkiewicz et al. ‘02] Artec Group [Newcombe et al. ’11] KinectFusion
Handling Deformations [Chang & Zwicker ’11] [Brown & Rusinkiewicz ’07] [Li et al. ’09]
Using Human Body Priors [Cui et al. ‘12] [Weiss et al. ’11] [Tong et al. ‘12]
Challenges deformation, clothing & props daily environment low cost
Global Non-Rigid Registration 65
3D Scanning
Automatic Reconstruction
3D Printing
http://cs621.hao-li.com Thanks! 69
Recommend
More recommend
