Gianfranco Doretto / Research / Project

3D Object Modeling

Range data registration in the frequency domain

Description

In this project we study the problem of aligning 3D range data in the frequency domain, and propose a technique that could be used for unsupervised registration in 3D object modeling.
Three-dimensional models of rigid objects can be obtained in two ways: from 2D imagery or from 3D range data. In the first case, the geometry of the surface of the object is typically recovered by matching sets of features extracted from images of the object taken from different viewpoints. In the second case, the depth information is provided by a range camera and the shape of the object is reconstructed by aligning sets of range data containing partially overlapping patches of surface.
While most of the registration techniques perform the alignment iteratively in the space domain, if we consider the corresponding problem in the frequency domain it is possible to decouple the estimation of the rotation from the estimation of the translation. This leads to an efficient algorithm that estimates motion parameters in closed-form and that can be used for 3D object modeling. The construction of accurate 3D models has several applications such as industrial prototyping, medical data visualization, cultural heritage preservation, etc.
The main contributions of our approach are:
  • We pose the 3D alignment problem in the frequency domain where the estimation of the motion, that has six degrees of freedom, can be cut down to two estimation problems of three degrees of freedom.
  • We provide a closed-form solution to the alignment problem that extends to 3D the previous 2D frequency domain alignment techniques.
  • We discuss the implementation of the algorithm and give several practical provisions that significantly improve the robustness of the estimation.
  • We extensively analyze the performance of the algorithm, and show that the estimation error remains within 3 degrees with probability of 90%.
  • We use the algorithm in the 3D object modeling pipeline, and build several examples of 3D models.

Results

Examples of 3D models reconstructed from partial views.

Bas-relief by Donatello

15th century, altar of Saint Anthony's Church, Padua, Italy.

Holy Virgin with Child by Giovanni Pisano

13th century, Arena's Chapel, Padua, Italy

Apollo

Medusa

Turbine

Vertebra

Related publications

  • Lucchese, L., Doretto, G., and Cortelazzo, G. M.
    A frequency domain technique for range data registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(11):1468–1484, November 2002.
    Details   BibTeX   PDF (3.3MB )  
  • Cortelazzo, G. M., Doretto, G., Lucchese, L., and Totaro, S.
    Three dimensional view registration by a frequency domain technique. In Proceedings of the European Signal Processing Conference, pp. 1289–1292, Island of Rhodes, Greece, September 1998.
    Details   BibTeX   PDF (555.2kB )  
  • Cortelazzo, G. M., Doretto, G., Lucchese, L., and Totaro, S.
    A frequency domain method for registration of range data. In Proceedings of the IEEE International Symposium on Circuits and Systems, pp. 518–521, Monterey, CA, USA, May 31--June 3, 1998.
    Details   BibTeX   PDF (468.7kB )  
  • Lucchese, L., Doretto, G., and Cortelazzo, G. M.
    Frequency domain estimation of 3D rigid motion based on range and intensity data. In Proceedings of the International Conference on 3D Digital Imaging and Modeling, pp. 107–112, Ottawa, Ontario, Canada, May 1997.
    Oral Presentation
    Details   BibTeX   PDF (489.7kB )  

Jamp to