Deep recursive segmentation networks

Abstract

This paper considers building recursive layers that are unrolled to increase depth of the architecture in the task of semantic segmentation. Currently, results from the state-of-the-art models illustrate that increasing the physical depth, that is adding more layers with skip connections, effectively boosts the segmentation performance. However, there are two main issues for the very deep models. Firstly, deeper models requires more labeled data to train, which is labor expensive for computer vision tasks, such as detection and segmentation. Secondly, most of very deep models are feed forward, which fails to support rapid visual recognition. In this work, Deep Recursive Segmentation Networks (DRSN) are proposed that reuse a portion of parameters during feed forward process. DRSN have two exceptional advantages over previous models. Firstly, DRSN contains a limited amount of parameters meanwhile making deeper model depth. Secondly, DRSN support rapid visual recognition that is vital in some applications, such as robots and autonomous cars. While utilizing only 15% of parameters of previous FCN models, we achieve 80% the performance on the pixel accuracy.