Evidential fully convolutional network for semantic segmentation

We propose a hybrid architecture composed of fully convolutional network (FCN) and Dempster-Shafer layer for image semantic segmentation. In the so-called evidential FCN (E-FCN), an encoder-decoder first extracts pixel-wise feature maps from input image. A then computes mass functions at each pixel location based on distances to prototypes. Finally, utility performs segmentation allows imprecise classification ambiguous pixels outliers. end-to-end learning strategy jointly updating parameters, which can make use soft (imprecise) labels. Experiments using three databases (Pascal VOC 2011, MIT-scene Parsing SIFT Flow) show that proposed combination improves accuracy calibration by assigning confusing multi-class sets.