Convolutional Random Walk Networks for Semantic Image Segmentation

TitleConvolutional Random Walk Networks for Semantic Image Segmentation
Publication TypeMiscellaneous
Year of Publication2016
AuthorsBertasius, G., Torresani L., Yu S. X., & Shi J.
Keywordsrandom walk, semantic segmentation, spectral graph partitioning

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, the use of large receptive fields and many pooling layers, cause blurring and low spatial resolution inside the deep layers, which often lead to spatially fragmented FCN predictions. In this work, we address this problem by introducing Convolutional Random Walk Networks (RWNs) that combine the strengths of FCNs and random walk based methods. Our proposed RWN jointly optimizes pixelwise affinity and semantic segmentation learning objectives, and combines these two sources of information via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. We show that such a grouping mechanism improves the semantic segmentation accuracy when applied in the deep low spatial resolution FCN layers. Our proposed RWN fully integrates pixelwise affinity learning and the random walk process. This makes it possible to train the whole network in an end-to-end fashion with the standard back-propagation algorithm. Additionally, our RWN needs just 131 additional parameters compared to the state-of-the-art DeepLab network, and yet it produces an improvement of 1.5% according to the mean IOU evaluation metric on Pascal SBD dataset.

ICSI Research Group