On learning higher-order consistency potentials for multi-class pixel labeling
ECCV'12 Proceedings of the 12th European conference on Computer Vision - Volume Part II
Dynamic programming for approximate expansion algorithm
ECCV'12 Proceedings of the 12th European conference on Computer Vision - Volume Part III
Computer Vision and Image Understanding
| Hi-index | 0.00 |
This paper presents a new adaptive graph-cut based move-making algorithm for energy minimization. Traditional move-making algorithms such as Expansion and Swap operate by searching for better solutions in some predefined moves spaces around the current solution. In contrast, our algorithm uses the primal-dual interpretation of the Expansion-move algorithm to adaptively compute the best move-space to search over. At each step, it tries to greedily find the move-space that will lead to biggest decrease in the primal-dual gap. We test different variants of our algorithm on a variety of image labelling problems such as object segmentation and stereo. Experimental results show that our adaptive strategy significantly outperforms the conventional Expansion-move algorithm, in some cases cutting the runtime by 50%.