Affine-invariant shape matching and recognition under partial occlusion

In this paper, a new approach for matching and recognizing affine-distorted planar shapes is proposed, which allows for partial occlusions. In our approach, each shape is divided into a sequence of ordered affine-invariant segments based on the properties of curvature scale space (CSS) shape descriptor. Then Smith-Waterman algorithm is applied for matching the pair of affine-invariant segment sequences. Finally the point correspondences along the two shapes are obtained and the affine transform is estimated accordingly. Experiments show that our algorithm is effective for matching and recognizing shapes which are distorted by affine transforms, including translation, rotation, scaling, and shearing. Furthermore, it is capable of dealing with the case where one of the shapes or both shapes are partially occluded.