Selective Symmetry Detection for Precise Blob Identification

The points generated by the interest map I(p) suffer from reduced accuracy due to the simplifications Sela proposes. Furthermore, these points return the center of regions of interest and not the contours or structure of the regions. We have investigated the use of the symmetry line data to reconstruct the ``blob'' detected at p. The lines of symmetry at multiple scales and orientation do contain information about the structure of the region of symmetric enclosure. However, due to the many simplifications imposed upon the calculation, the data involved are compressed and the computation of this ``inverse symmetry transformation'' does not possess the accuracy needed to properly segment the ``blob'' for our purposes. A more accurate description of the ``blob'' would require a higher quality version of the symmetry transform.

Consequently, we propose the use of a high quality, but slower, symmetry transform as a post-processing stage to the previously described real-time interest operator. The significant points generated by the interest map I(p) serve as points of attention for this higher quality analysis. However, we do not attempt to actually invert the symmetry lines generated by this calculation to approximate the blob's contour. Instead, the annular sampling regions used to detect symmetry are deformed to function as templates for the specific shapes to be identified. This technique is similar to template matching. However, it utilizes the principles of symmetric enclosure to detect the desired blob and not merely the intersection of a template with edge data. This restricts the false alarms that might trigger simple template matching (as shown later in Figure ). The following derivations outline the development of a symmetric enclosure measure for deformable annular sampling regions.