410 likes | 447 Views
The GeoTopotransform model proposes a deformation-driven approach for shape correspondence, supporting topological changes without a fixed number of segments. The method efficiently computes continuous part correspondences for diverse man-made shapes, assessing shape matching quality based on structure preservation.
E N D
Deformation-Driven Topology-Varying 3D Shape Correspondence Ibraheem Alhashim Kai Xu YixinZhuang Junjie Cao Patricio Simari HaoZhang Presenter: Ibraheem Alhashim Simon Fraser University
Shape Correspondence • Fundamental task in: Classification Shape morphing Statistical shape modeling Object recognition
Shape Correspondence • Corresponding man-made 3D shapes is challenging • Large variability in geometry & structure • Real-world data is inconsistent & unlabeled
Shape Correspondence • Part-level correspondences [Xu 12] [Zheng 13] [Kalogerakis 12] [Jain 12] [Averkiou 14]
Shape Correspondence Continuous fine-grained correspondence is critical for continuousshape blending [Alhashim et al. 14]
Previous Works • Rigid alignment not sufficient for diverse shapes [Golovinskiy & Funkhouser 08]
Previous Works • Co-analysis methods • Coarse results • Forced correspondence • Set of shapes [Zheng et al. 14] [Kim et al. 13] [Huang et al. 14] [Laga et al. 14]
Deformation-Driven Shape Matching • Best matching = minimal self-distortion as we deform one shape to match the other [Sederberg & Greenwood 92] [Zhang et al. 08]
Challenge How to apply a deformation-driven search to complex man-made shapes? • Many disconnected components • Semantically similar yet very different • Discrepancy in part count & structural relations Back Seat Legs
Our Proposal The GeoTopotransform • Piece-wise continuous part correspondence • Supports topological changes • No prior or fixed number of segments • Efficient to compute • Works on pairs
Our Proposal The GeoTopotransform Distortion Energy Deformation model
Deformation Model • Deformation suitable for man-made shapes • Supports disconnected components • Structure-aware (preserving part relations) • Allows for topological changes
Self-Distortion Energy • Structural distortion in three terms: • Distortion on all pairs of connected parts • Connectivity between parts • Solidity measure for parts changing topology
Shape Representation • A structure graph of part skeletons [Alhashim et al. 2014] • Skeletons are fitted by parametric curves / sheets Parametric curves Parametric sheet
Structural Rods 3D shape Curve-sheet abstractions Structural rods
Energy • Distortion term Ed • Overall change of part arrangements • Change in angle
Energy • Distortion term Ed Best correspondence Before deformation After deformation
Energy • Connectivity term Ec • Relative length of shortest rods before and after deformation Deformed shape Source shape Target
Energy • Solidity term Es • Ratio between the volume of a part to the volume of its convex hull • Measures the effect of a split / merge Low High High
Energy • Solidity term Es ? ?
Deformation Process • Deform-to-fit matched parts, then propagate 1. Align centers 3. Deform towards target 2. Match extremities 4. Propagate edit to others Curves Sheets
Search • Search tree path: set of matched parts on the source • Beam search + pruning leg front-leg front seat-seat back-back leg back-leg back back bar-back bar 3D shapes Curve-sheet abstractions
Applications fully automatically! • Shape blending
Applications Shape Classification Topological medoid
Evaluation • Ground truth • 75 shapes, 5 categories (chair, airplane, table, bed, velocipede) • Fine and coarse labels
Evaluation • [Xu 12] Fuzzy part correspondence (baseline) • Works on pairs • Match based on part OBB similarity • [Zheng 14] Recurring part arrangements • Find semantic consistency between part arrangements • Performs better than co-segmentation in the presence of large shape variability • [Kim 13] Deformable part-based templates • Better suited for large shape sets • Supports poorly segmented inputs + can be fully auto.
Evaluation • Fine-grained correspondence benchmark
Evaluation • Coarse correspondence benchmark (co-analysis)
Summary • GeoTopo:topology-varying deformation model for a fine-grained correspondence search • Key contribution: a deformation model and a self-distortion energy, defined on structural rods, assess shape matching quality based on preservation of structure Our framework shows promising results on challenging datasets with much room for improvement
Limitations Initial segmentation Large geo. and topo. differences
Future Work • Segmentation • Online shape repositories are not well segmented • Incorporate a segmentation search along with the correspondence search High energy Low energy segmentation
Future Work • Co-analysis • Fine-grained correspondence on a set • Looking for consistent assignments
gruvi.cs.sfu.ca/project/geotopo THANK YOU! ACKNOWLEDGMENTS Anonymous reviewers, authors who provided code, funding from: NSFC
Automatic Segmentation Convex. Analysis SDF Con. Aware Approximate Convexity Analysis
Structure Preservation • Edit propagation • Reinforce symmetry and contact relations