Reviews: Submodular Field Grammars: Representation, Inference, and Application to Image Parsing

The key problem is that splitting the image into *arbitrarily-shaped* pixel regions to associate with the production rules is computationally difficult in general. This paper proposes to associate formal grammar production rules with submodular Markov random fields (MRF). The submodular structure of the associated MRF allows for fast inference for a single rule into arbitrarily-shaped subregions and a dynamic-programming-like algorithm for parsing the entire image structure. The experimental results show that the method is indeed much faster than previous methods. Pros: 1) Well-written and easy to read even though some of the details are fairly technical.