IEEE 2014 Conference on Computer Vision and Pattern Recognition Beam Search for Solving QP Beam Search for Solving QP Results Results Acknowledgment Acknowledgment Problem and Motivation Problem and Motivation Approach Approach 1. 1. Extracting superpixels Extracting superpixels 2. 2. Incorporating mutex in the standard CRF Incorporating mutex in the standard CRF formulation formulation 3. 3. Formulating CRF inference as QP Formulating CRF inference as QP 4. 4. Beam search for solving QP Beam search for solving QP 5. 5. Learning – piecewise Learning – piecewise How to Specify CRF Energy? How to Specify CRF Energy? CRF Inference as QP CRF Inference as QP Specifying Mutex Constraints Specifying Mutex Constraints Scene Labeling Using Beam Search Under Mutex Constraints Anirban Roy and Sinisa Todorovic Input Image Semantic segmentation with Mutex Semantic segmentation without Mutex Mutex violati ons Appearance features of the superpixels Smoothness and Contex t ∞ Pixelwise accuracy(%) can be arbitrary Assignment vector Matrix of CRF potentials NSF RI 1302700 MUTual EXclusion = (object, object, relationship) State: label assignment Heuristic function: Score: Superpixel Class label Method MSRC Test time Galleguillos et al. CVPR 10 70.4 N/A Gould et al. ICCV 09 76.4 N/A Payet et al. PAMI 12 82.9 30-32s Krahenbuhl et al. NIPS 12 86.0 0.2s Yao et al. CVPR 12 86.5 N/A Zhang et al. CVPR 12 87.0 N/A Ours 91.5 0.8s Method Stanford Backgrou nd Gould et al. ICCV 09 76.4 Munoz et al. ECCV 10 76.9 Singh et al. CVPR 13 74.1 Ours 81.1 Matrix of mutex constraints maximum score next state previous state must be