Recognition of Hand-drawn Sketches with Marcov Random Fields

Freehand sketches are complex for recognition. Individual fragments of the drawing are often ambiguous to be interpreted without contextual cues. Markov Random Field (MRF) that ends up with a global model by simply specifying local interactions can naturally suffice the requirement. In our project, a recognizer based on MRF has been constructed to jointly analyze local features in order to incorporate contextual cues during inference process. Shapes in sketches are detected and matched to a deformable template. Whereas standard belief propagation (BP) is not guaranteed to converge for inference on graphical models with loops, we find that loopy belief propagation (LBP) does converge in our experiment. The final recognition is found as the MAP marginal by global belief propagation.