Surface Correspondence Benchmark

← Back

Download the benchmark (184Mb)

Data Sets

SCAPE: 71 meshes representing a human body in different poses. All the meshes were fit to scanner data with a common template (thus corresponding vertices have same IDs).
TOSCA: 80 meshes representing people and animals in a variety of poses. The meshes appear in 8 groups with common topology (corresponding within the same class have the same ID).
Watertight: 400 meshes arranged evenly in 20 object categories, many of which are articulated figures (humans, octopus, four-legged animals, ants, etc.). We selected 11 classes for our experiments that have well defined correspondences and genus zero. In addition, we excluded two human models with non-zero genus.

Download the benchmark (184Mb): this package includes meshes from aforementioned datasets (in *.off format, with consistent normal orientation and after delaunay triangulation (via edge flips) filter. It also contains manually selected ground truth correspondences and some useful processing scripts.

If you use the datasets you should cite:

[1] The correlated correspondence algorithm for unsupervised registration of nonrigid surfaces
     Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Hoi-Cheung Pang, and James Davis.
     Proc. Neural Information Processing Systems (NIPS), 2004

[2] The correlated correspondence algorithm for unsupervised registration of nonrigid surfaces
     Alex Bronstein, Michael Bronstein, and Ron Kimmel
     Springer, 2008

[3] SHREC: shape retrieval contest: Watertight models track
     Daniela Giorgi, Silvia Biasotti, Laura Paraboschi, 2007

[4] Blended Intrinsic Maps
     Vladimir G. Kim, Yaron Lipman, and Thomas Funkhouser
     SIGGRAPH 2011

Evaluation Metric

For some datasets we have ground truth vertex-to-vertex correspondence (depicted by same color of vertices in SCAPE and TOSCA). If it is not availiable manual coarse correspondences are defined by a user (see coarse points in SCAPE, TOSCA, and Watertight):

To evalulate performace of an algorithm we map every vertex that has a ground truth correspondence. A geodesic distance between the predicted correspondence and the true correspondence is recorded (see isocontours starting from a blue point).
Note that some methods are often confused by bilateral reflective symmetry (for instance, they often map left side to right side in human and animal classes). To distinguish between this error and all other errors we also evaluate method's performance on a ground truth that accept such symmetric flips (cumulative per mesh error calculated with respect to two ground truths, and the map with the smallest cumulative error is chosen).


We compared performance of Blended Intrinsic Maps [1], Möbius Voting [2], Generalized Multi-Dimensional Scaling [3] and Heat Kernel Maps [4] and Deformation-Driven [5] to detect inter-surface correspondence. You can see the results here (222Mb).

For individual methods, please cite:

[1] Blended Intrinsic Maps
     Vladimir G. Kim, Yaron Lipman, and Thomas Funkhouser
     SIGGRAPH 2011

[2] Möbius Voting for Surface Correspondence
     Yaron Lipman and Thomas Funkhouser
     SIGGRAPH 2009

[3] Generalized multidimensional scaling: a framework for isometry-invariant partial surface matching
     Alex Bronstein, Michael Bronstein, and Ron Kimmel
     Proc. National Academy of Sciences (PNAS), 2006

[4] One point isometric matching with the heat kernel
     Maks Ovsjanikov, Quentin Merigot, Facundo Memoli, and Leonidas Guibas
     SGP 2010 (Symposium on Geometry Processing)

[5] Deformation-driven shape correspondence
     Hao Zhang, Alla Sheffer, Daniel Cohen-Or, Qingnan Zhou, Oliver van Kaick, and Andrea Taglisacchi
     SGP 2008 (Symposium on Geometry Processing)