There are two related projects:

1. Knowledge-guided object segmentation
2. Estimating the spatial extent of an object

1. Goal: Segmenting specified objects from an arbitrary background by combining a knowledge-guided segmentation method with an object model. However, the model might not be ideal for the task of segmentation (see below left image).

Therefore, we "tailor" the model to fit the task using feedback from segmentation of a labeled training set. The idea is to modify the model such that, in the next iteration, regions A are included in the segmentation and regions B are excluded.

2. Model modification comprises adding, removing, or reweighting Gaussians in the GMM-based object model. For example...

Segmentation leaves a hole in the "golf course" object By adding a Gaussian to the model, we can fill this hole.

Note: Actually, a rotation-invariant version of the original texture motif-based object model is used in this work.

3. The process stops when a desired segmentation quality is reached. The criterion could be user-defined, say percentage deviation from a segmentation mask.

1. Suppose we know the point location of an object and we need to estimate its spatial extent. This is often the case for objects in geographical databases where only the latitude and longitude are given.

2. Based on the texture motif model, we can assign motif labels to the pixels in the object. We model the spatial arrangement of the motifs using a multi-level logistic (MLL) Markov random field (MRF).

A "golf course" object		Motif label assignment

3. Starting from a seed point inside a new object, a bounding box is iteratively expanded in the direction that is most similar to the MLL model.

3. The example results below illustrate the iterative growth of the bounding box. The numbers of iterations are shown at various stages.

B. Sumengen, S. Bhagavathy, and B. S. Manjunath, "Graph Partitioning Active Contours for Knowledge-Based Geo-Spatial Segmentation," CVPR Workshop on Perceptual Organization in Computer Vision (POCV), Washington DC, Jun 2004. Abstract

S. Newsam, S. Bhagavathy, and B. S. Manjunath, "Object Localization Using Texture Motifs and Markov Random Fields," International Conference on Image Processing, Barcelona, Spain, Sep 14-17, 2003. Abstract