acelricol

If you want do delete your site from our spam bases - just email us with domain of your site:

abuse-here@inbox.ru

thank you!

The shape context paper presents '3D object recognition' results on a collection of 2D images that contain only one instance of an object with no background clutter in the image. I haven't seen anybody use a dataset like this in a few years.

The paper by Belongie and Malik presents a robust method for shape matching and object recognition by using shape contexts. They compute 1-1 correspondences between shapes and use them for shape matching. As far as object recognition is concerned, the method seems to do fairly well. However, I donot advocate the use of shape contexts (SC) for Digit or Handwriting Recognition (HWR). For HWR there are far better and simplistic methods for recognition. For example, Dynamic Time Warping (DTW) is more popularly used for HWR. A quick comparison reveals that DTW is a far more simpler technique to implement than SC. On brief look at the Fig 8 (failed cases) in the Belongie paper, it appears to me that SC does almost as bad or worse than DTW for difficult cases. Also, one of the major complaints with DTW is that it is quadratic and lot of efforts have been taken to linearize it. On comparison, the SC technique is inherently cubic (albeit near linear claims in the paper) on feature points and hence is not applicable for real-time HWR. For a preliminary comparison of SC, DTW and related techniques on recognizing George Washington's handwriting see: Using Corner Feature Correspondences to Rank Word Images by Similarity by Jamie L. Rothfeder, Shaolei Feng and Toni M. Rath.
SC is worse.

Please discuss Shape Matching and Object Recognition Using Shape Contexts