In this work, we used a bag-of-words (BoW) representation for glyph based on patch-based 2-ring HOOSC descriptors computed at sample glyph positions (we tested 4 different spatial context sc for this local descriptors; sc = 1/1 corresponds to 128x128 patch for computing the feature). As for the vizualization, we relied on the t-distributed Stochastic Neighborhood Embedding (t-SNE) methodology applied to these BoW representations.
Please click on the links to see visualization of the 10-class glyph samples from stone monuments.
HOOSC with sc = 1/1 (Most global shape descriptor) |
HOOSC with sc = 1/2 |
HOOSC with sc = 1/4 |
HOOSC with sc = 1/8 (Most local shape descriptor) |