Skull Assembly and Completion using Template-based Surface Matching Li Wei, Wei Yu, Maoqing Li 1 Xin Li* 2 1 School of Information Science and Technology Xiamen University 2 Department of Electrical and Computer Engineering Center for Computation and Technology Louisiana State University *EMail: xinli@lsu.edu 3DIMPVT 2011 1 / 17
Introduction Goal To assemble and repair fragmented skulls to complete models. Motivation Skull Completion is important in forensic and anthropological facial reconstruction. The traditional state-of-the-art approach: Time consuming Domain expertise required Subjective when large regions are missing Proposed Approach: Computer-aided completion in digital environment 2 / 17
Introduction Assembling Completion Results Conclusion Related Work 3 / 17
Algorithm Pipeline 4 / 17
Fragmented Skull Assembling In: A set of fragments M i and template S. Out: A set of rigid transformations T i (applied on M i ), so that the arrangement of all fragments in world coordinates well approximates S. 5 / 17
Feature Exaction Feature Extraction Slippage Feature on each Fragment M i. Signature Spin Image [Johnson et al. 1999]. Why Spin Image? Comparing the Geometric Properties of regions. Our Modified Spin-Image Signature Resolution-Invariance Direction Filtering 6 / 17
Finding Coarse Correspondence I/O In: Spin image signatures on fragments feature points and template points. Out: Correspondence from feature points (on fragments) to the template. The most possible many-to-many mapping (most similar corresponding pairs of points) is the coarse correspondence. 7 / 17
Fragment-Template Matching Optimal Correspondence Exact a most isometric sub-set [Tevs et al. 2009]. Local Registration Compute rigid transform by optimal corresponding. 8 / 17
Inter-Fragment Refinement Avoid Intersection Between Fragments Refinement: Taking the computed transformations as starting variables, to minimize the objective function composed of three terms: E(T 1,..., T n) = αe I + βe F + γe T, where E I = i,j n,i<j Int(T i(m i ), T j (M j )), E F = i v T i (M i ) Dis(v, S) 2, E T = i S(T i). 9 / 17
Skull Completion I/O In: The assembled skull with damage regions, and a template. Out: A repaired skull The completion is based on the surface matching between the the template and subject skull. 10 / 17
Template-based Completion Non-rigid Template/Subject Mapping In: A template model, and the assembled skull. Out: Affine transformations on each template vertex Minimizing the following matching energy: E = λ 1 E data + λ 2 E shape where E data = Dis 2 (T i + v i, M) v i S E shape = T i T j 2 e i,j,v i,v j S 11 / 17
Template-based Completion I/O In: Deformed template and assembled object skull. Out: Repaired object skull Cut and paste. Filling small holes/gaps[liepa et al. 2003] Smoothing[Kobbelt et al. 1998] 12 / 17
Experimental Results 13 / 17
Completion in Various Cases 14 / 17
Conclusion Skull Assembling: Fragment-template matching > Spatial relationship of fragments. Robust and efficient partial matching by improved spin image. Global refinement to reduce self intersection. Skull Completion: Non-rigid registration from template to object. Robust hole-filling and smoothing algorithm. 15 / 17
Acknowledgements The Skull Data are provided by the Forensic Anthropology and Computer Enhancement Services (FACES) Laboratory at Louisiana State University This work is supported by Louisiana Board of Regents Research Competitiveness Subprogram (RCS) LEQSF(2009-12)-RD-A-06, PFund: NSF(2009)-PFund-133, and LSU Faculty Research Grant 2010. Project Website: http://www.ece.lsu.edu/xinli/skullassembly.html 16 / 17
On-going and Future Work Assembly Increase accuracy by analyzing break curves. Completion Intergrate the completion using symmetry. Facial reconstruction Digital face modeling on completed skull 17 / 17