This paper provides an approach to measure the similarity (distance) between two shapes. They uses several points sampled from the contours to represent the shape. Each sample point has a descriptor to describe the coarse distribution of the rest parts of the shape by using histogram of the angle and distance of other sample points. From the point descriptors, the problem of finding the correspondence between two shapes becomes the problem of finding for each sample point on another shape that has the most similar shape context.
After obtaining the correspondences between points on two shapes, an aligning transformation (warping) from one shape to the other is estimated. The magnitude of the aligning transform is viewed as a measure of shape similarity. The overall similarity between the two shapes is computed as the sum of the matching errors between corresponding points and the magnitude of the aligning transform.
K-medoids (a variant version of K-means) is adopted to select the prototypes that should be stored, and a K-NN classifier is used in the recognition job. By making experiments on 3D object database, they have shown that their algorithm is invariant to several image transformations.
This paper provides a method to evaluate the distance between two shapes in a very dissimilar way compared with previous works. By using shape context descriptor, they do not require any key-point (e.g., maxima of curvature or inflection points) or the grayscale values inside the silhouette to describe the shape.
After obtaining the correspondences between points on two shapes, an aligning transformation (warping) from one shape to the other is estimated. The magnitude of the aligning transform is viewed as a measure of shape similarity. The overall similarity between the two shapes is computed as the sum of the matching errors between corresponding points and the magnitude of the aligning transform.
K-medoids (a variant version of K-means) is adopted to select the prototypes that should be stored, and a K-NN classifier is used in the recognition job. By making experiments on 3D object database, they have shown that their algorithm is invariant to several image transformations.
This paper provides a method to evaluate the distance between two shapes in a very dissimilar way compared with previous works. By using shape context descriptor, they do not require any key-point (e.g., maxima of curvature or inflection points) or the grayscale values inside the silhouette to describe the shape.
Reference:
S. Belongie, J. Malik and J. Puzicha, "Shape Matching and Object Recognition Using Shape Contexts." IEEE Trans. Pattern Anal. Mach. Intell, 24(4), pp. 509-522, 2002.
沒有留言:
張貼留言