Joint recognition / segmentation with cascaded multi-level feature classification and confidence propagation

In this paper, we propose a semantic object segmentation method based on cascaded superpixel-wise classification and segment-wise object class ranking. Inspired by the overwhelming parsing ability of human's visual system in which non-local information are widely taken into consideration, our approach refers to higher-order information in case of ambiguous classifications. Different from many works on structured prediction for scene understanding, our work does not use complicated global probabilistic model, but adopts hierarchical cascaded classification for different levels of features. Another contribution is the confidence propagation through segment-wise object class ranking. Unlike many existing works which treat each classification unit equally, our method automatically discovers confident classifications and passes confidence to uncertain areas within segments obtained by hierarchical image segmentation. Such label correction process can significantly boost the segmentation accuracy.