| （英） |
Pixel labeling is one of the most classical and important problems in the field of computer vision because it has a variety of applications. We tackle two major challenges of pixel labeling: (i) how to deal with the large solution space, and (ii) how to learn the relationships between neighbor labels effectively. For the first challenge, we present two neighbor-aware fast optimization methods. One is the fast optimization method for general pixel-labeling problems based on Markov random field (MRF) models where the smoothness between the neighbor labels is forced. The other is the fast optimization method for the special case of pixel-labeling problems where the neighbor labels are forced to be connected. For the second challenge, we present two novel neighbor-aware learning methods that boost the performance of pixel labeling. Based on the mathematical relationship between the fixed point iteration of dense conditional random field (CRF) and recurrent convolution, we present a new model based on dense CRF, which automatically learns the relationships between neighbor labels from training data and enables joint training with deep neural networks. In addition, we present a novel problem setting (pixelRL), and an effective neighbor-aware learning method for pixelRL named reward map convolution. PixelRL is a novel pixel-labeling problem combined with reinforcement learning, where the label is a sequence of actions at each pixel, and its objective is to maximize the accumulated total rewards at all pixels. |