Identification of Partially Visible Variational Shape Classes

Variational shape classes. Top: 2% vertex variation. Bottom edge length and angle variation (5,10%). High variation contours are particularly difficult to recognize when they are partially overlapping.

Many shapes are not exact, and can vary to differing degrees. A shape class captures this extent of variability. The main contribution of this thesis is formalising a theory of shape class modelling. We provide two methods of deformational shape class modelling. We use a novel polygonal representation scheme with local invariant relations of edge vectors with supersegments. This is translation, rotation and scale invariant which saves us from a searching for appropriate transformations for the candidate models. This model of representation is can be made very stable under varia- tions using higher length supersegments.

We venture with three algorithms for identification of variational shape classes with supersegment matching strategy:

1. Probabilistic Directed Hausdorff Distance with Discretized Error Interval, which is much more efficient and robust than existing hausdroff distance based algorithms and also can handle cases of occlusion.
2. GA, where we extend the earlier attempts of shape matching with genetic algorithms for variational shape classes.
3. Maximal Supersegment Model, which is a novel robust and efficient algorithm for variational shape class recognition.

As the title portrays, all the three algorithms are capable to identify partially visible shape classes from a given image.

Occluding Image Contours. Two animal shapes overlapping by increasing degrees. All three cases are recognized by the maximal supersegment model.

Goswami, Anjan; and Amitabha Mukerjee; 1999. 
    Modeling and recognition of Planar Shape Classes, 
    Proceedings Fourth Intl Conf on Advances in Pattern Recognition
	and Digital Technology, Calcutta, December 27-29,
	1999. [ gzipped Postscript, (82696 bytes) ]

Goswami, Anjan; and Amitabha Mukerjee; 2000
    Multi-Threshold Hausdorff Distance with Maximal Arc-Length
        Matching for Variational Shape Recognition,
    ECCV-2000, Dublin Ireland, June 2000. (under review). [ gzipped Postscript (82 kB) ].

[Thesis List] [ME Home Page]