Thomas Sebastian

Research Projects

Medical Image Segmentation using Skeletally Coupled Deformable Models

Segmentation uses a deformable model that allows for region competition (for details see skeletally coupled deformable models). The main application domain is segmenting carpal bones from CT images.

Matching curves

Finds an optimal correspondence (alignment) between two curves. An alignment curve is introduced to ensure symmetric treatment of both curves. Dynamic programming is used to find the optimal correspondence. The application include handwritten character recognition, prototype formation, shape morphing, and object recognition. Details.

Computational Atlases

Curve matching framework is used to compute an average of a set of curves. This average can serve as the mean of a collection of anatomical structures and can be used for distinguishing diseased from normal states. Details

Shock graph matching for recognizing shapes

The similarity of shapes is defined in terms of the minimum extent of deformation necessary to morph one shape to another. The space of shapes and deformation paths are discretized. An edit-distance algorithm which finds the optimal deformation path in polynomial time. The approach is robust to a variety of visual transformations including boundary perturbations, articulation and deformation of parts, viewpoint variation, segmentation errors and partial occlusion. 100% recognition rate is achieved on two distinct databases of 99 and 216 shapes each. Details

Indexing into large shape database

We have created a database consisting of 1032 shapes, developed a coarse-scale version of shock graph matching and an exemplar-based indexing scheme, to make shock-based indexing into large shape databases practical. Details

Online Publications

Segmentation of Carpal Bones from 3D CT Images using Skeletally Coupled Deformable Models. MICCAI 98. Abstract. Postscript
Constructing 2D Curve Atlases. MMBIA2000. Abstract. Postscript PDF (Copyright IEEE)
2D-3D Registration Based on Shape Matching. MMBIA2000. Abstract. Postscript (Copyright IEEE)
Shape matching using edit-distance: an implementation. SODA2001. Abstract. Postscript PDF
Alignment-based Recognition of Shape Outlines. IWVF2001. Abstract. Postscript PDF (Copyright Springer Verlag)
Recognition of Shapes by Editing their Shock Graphs ICCV2001.Abstract. Postcript, PDF (Copyright IEEE)
Curves vs Skeletons in object recognition. ICIP2001. Abstract. Postscript PDF (Copyright IEEE)
Shock-based Indexing into Large Shape Databases. ECCV2002. Abstract. Postscript , PDF (Copyright Springer Verlag)

Online Presentations

Symmetry-based segmentation and recognition. PPT.

Shape-based Object Recognition. PPT.

Brown University
Laboratory for Engineering Man/Machine Systems (LEMS)
Box D, Division of Engineering
Providence, RI 02912
Tel: 401-863-2177
tbs@lems.brown.edu