Contour Grouping for Generic Object Modeling and Recognition

Robert Bergevin


Abstract

In this talk, I present a few image analysis algorithms developed in our group in the past years. Their connecting theme is qualitative shape modeling and recognition of multi-part 3D objects from single static images.

Contour-based perceptual grouping is used in all cases. We start with the multi-scale description of 2D contours using constant-curvature primitives. The main ideas behind the MuscaGrip algorithm are shown, various comparative results from our first implementation are presented and efforts to improve the performance are discussed. Application of the algorithm to the extraction of generic solid parts in static images is also discussed. Extraction of a structured contour map from a static image is presented next. Using the ConStruct algorithm, image junctions are detected, characterized and paired in order to build a fine-scale contour description adapted to the complexity of the object's shape. A two-stage process is used in order to build each contour description. Finally, the PLASTIQUE system is introduced. Foreground objects are detected and modeled using the spatial structure of approximated parts. Parts are categorized using a fuzzy classifier and subsets of connected parts are used as an index into a database of descriptions from various images. Images of similarly-shaped objects observed in different contexts are retrieved from the database.


Maintained by Philippos Mordohai