Semi-local affine Parts for Object Recognition_object

A computational hypotheses

1. Extraction of elliptical affine plane.
2. Two kinds of descriptors are used to extract the characteristics of the elliptical affine plane in 2 images.
3. Based on the recent principles of Euclidean distance between two features to look for possible matches
4. The other two elliptical planes of an ellipse plane and its adjacent area are composed of triple. (What are the two other elliptical planes?)? If both pictures randomly take the following affine, it would be troublesome. )
5. According to the first ellipse plane, find the ellipse plane that it may match in another picture.
6. In triple, the Triple center affine map of one of the graphs is mapped to a possible matching triple in another picture, with the shape, size, and angle of the 3 ellipses compared.
7. After finding the matching triple, the other elliptic affine planes in its neighborhood are added to the matching region by greedy algorithm. For each addition, the affine matrix of the matching region center is estimated based on the least squares method. Until all possible matching ellipses have a characteristic distance of less than the threshold value.
8. Removal of matched regions containing less than 6-8 ellipses. (description does not match)
9. Merging the matching regions with the overlap part greater than a certain degree. As hypotheses.

The second converts the hypotheses into a parts that can be described completely by affine transformation.

The nearest of the two groups, from the center of the ellipse, the point of affine transformation is fully satisfied, according to the two sets of corresponding points, the affine transformation matrix and corresponding coordinate system are computed, and then two sets of ellipse sets are mapped into the coordinate system, then the average (descriptor, length, axis and direction of long axis) is taken as a parts

Thirdly, the parts of the object can be matched (trained) to eliminate the wrong hypotheses by using the above obtained.

The above parts is transformed into a confirmation image by affine transformation, as a hypotheses.
It then detects the number of areas contained in each hypotheses in each confirmation image.
Adds the number of all the elliptical regions contained in the same hypotheses (in all the confirmation images) and then sorts the parts by this value (rank), discarding the sorted parts.