Perceiving Systems, Computer Vision

Shape and pose-invariant correspondences using probabilistic geodesic surface embedding

2011

Conference Paper

ps


Correspondence between non-rigid deformable 3D objects provides a foundation for object matching and retrieval, recognition, and 3D alignment. Establishing 3D correspondence is challenging when there are non-rigid deformations or articulations between instances of a class. We present a method for automatically finding such correspondences that deals with significant variations in pose, shape and resolution between pairs of objects.We represent objects as triangular meshes and consider normalized geodesic distances as representing their intrinsic characteristics. Geodesic distances are invariant to pose variations and nearly invariant to shape variations when properly normalized. The proposed method registers two objects by optimizing a joint probabilistic model over a subset of vertex pairs between the objects. The model enforces preservation of geodesic distances between corresponding vertex pairs and inference is performed using loopy belief propagation in a hierarchical scheme. Additionally our method prefers solutions in which local shape information is consistent at matching vertices. We quantitatively evaluate our method and show that is is more accurate than a state of the art method.

Author(s): Tsoli, A. and Black, M. J.
Book Title: 33rd Annual Symposium of the German Association for Pattern Recognition (DAGM)
Volume: 6835
Pages: 256--265
Year: 2011

Series: Lecture Notes in Computer Science
Editors: Mester, Rudolf and Felsberg, Michael
Publisher: Springer

Department(s): Perceiving Systems
Research Project(s): 3D Mesh Registration
Bibtex Type: Conference Paper (inproceedings)
Paper Type: Conference

Links: pdf
Attachments: talk

BibTex

@inproceedings{Tsoli:DAGM:11,
  title = {Shape and pose-invariant correspondences using probabilistic geodesic surface embedding},
  author = {Tsoli, A. and Black, M. J.},
  booktitle = {33rd Annual Symposium of the German Association for Pattern Recognition (DAGM)},
  volume = {6835},
  pages = {256--265},
  series = {Lecture Notes in Computer Science},
  editors = {Mester, Rudolf and Felsberg, Michael},
  publisher = {Springer},
  year = {2011},
  doi = {}
}