A scene parsing method based on super-pixel and mid-level feature

Scene parsing can be formulated as a labelling problem that tries to label each pixel in an image with category of the object it belongs to, which involves the simultaneous detection, segmentation and recognition of all the objects in the image. A three stages method based on super-pixel and mid-level feature is proposed in this paper. First, super-pixels of the image are obtained by quick-shift. Second, the mid-level of each super-pixel are collected by aggregating the sift features in the super-pixel and its neighbor with sparse coding and max-pooling. Third, through CRF Models, which imposes consistency and coherency between labels, the globally optimal labeling results are obtained. Experimental results show that our method gains higher accuracy than previous methods.