ReSeg: A Recurrent Neural Network for Object Segmentation

摘要: We propose a structured prediction architecture for images centered around deep recurrent neural networks. The proposed network, called ReSeg, is based on the recently introduced ReNet model object classification. modify and extend it to perform segmentation, noting that avoidance of pooling can greatly simplify pixel-wise tasks images. ReSeg layer composed four networks sweep image horizontally vertically in both directions, along with final expands back original size. combines multiple layers several possible input as well which size, making suitable variety tasks. evaluate specific task segmentation three widely-used datasets, namely Weizmann Horse, Fashionista Oxford Flower. results suggest challenge state art may have further applications at large.