/SGA18_orientation_BCICP_dataset

Datasets tested in the paper "Continuous and Orientation-preserving Correspondence via Functional Maps"

Primary LanguageMatlab

Datasets

Here are the datasets tested in the paper "Continuous and Orientation-preserving Correspondence via Functional Maps" by Jing Ren, Adrien Poulenard, Peter Wonka and Maks Ovsjanikov.

Original Datasets can be found at:

In our setting, we remeshed each shape independently using LRVD algorithm to avoid overfitting. The remeshed meshes have ~5k vertices.

For each of the dataset (FAUST/TOSCA_Isometric/TOSCA_nonIsometric)

  • test_pairs.txt: each row shows a pair of shape names in this collection that we tested
  • vtx_5k/:
    • xxx.off: the remeshed shapes in this collection
    • corres/: for any pair (A.off, B.off) in this collection, "A.vts" stores the given ground-truth direct correspondence to "B.vtx", and "A.sym.vts" stores the ground-truth symmetric correspondence to "B.vtx"
    • segmentation/: for a pair (A.off, B.off), the segmentation of the two shapes are stored as "A_B.A.seg.sym" and "A_B.B.seg.sym". Moreover, "A_B.B.seg.nonsym" and "A_B.B.seg.nonsym" give the segmentation after breaking the symmetry. Please refer to Robust Structure-based Shape Correspondence for the details of the segmentation
    • maps/: stores the maps we computed and measured for the curves shown in the paper (with format "A_B.map", and the pairs are listed in the "test_pairs.txt")
      • BIM/: Blended Intrinsic Maps (note that the maps are 0-based)
      • PMF_hk/: Kernel Matching (note that the maps are computed on some samples of the mesh, and there are NaN indicating no correspondences)
      • WKSini_direct/: using WKS desriptors with our orientation-preserving term
      • WKSini_symm/: using WKS descriptors with out orientation-reversing term
      • WKSini_direct_BCICP/: WKSini_direct + BCICP
      • WKSini_symm_BCICP/: WKSini_symm + BCICP
      • SEGini_direct_BCICP/: using segmentations to compute the descriptors, then add the orientation-preserving operator to compute the initial functional map. Then add BCICP refinement.

Comments

  • The script run_example.m shows how to load the ground-truth correspondences, pre-computed segmentation, and saved maps. It is also included that how we measured the pre-vertex accuracy of the maps.
  • Our baselines are (1) BIM, (2) WKSini + ICP and (3) PMF. (2) is not included here, but the maps can be easily computed by applying 10 iterations of ICP to "WKSini_direct" and "WKSini_symm".
  • Measuring these maps on all the given correspondence should give the curves shown in the paper.
  • Please do not forget to cite the paper LRVD algorithm if you use the dataset in your work. Thanks (〃ﾉωﾉ)