Recurrent Segmentation for Variable Computational Budgets
State-of-the-art systems for semantic image segmentation utilize feed-forward pipelines with fixed computational costs. Building an image segmentation system that works across a range of computational budgets is challenging and time-intensive as new architectures must be designed and trained for every computational setting. To address this problem we develop a recurrent neural network that successively improves prediction quality with each iteration. Importantly, the RNN may be employed across a range of computational budgets by merely running the model for varying numbers of iterations. The RNN achieves results comparable to state-of-the-art systems in image segmentation on PASCAL VOC 2012 and Cityscapes segmentation datasets, however the RNN performs this task with a fraction of the computational budget. Finally, we demonstrate how one may exploit the properties of the RNN to efficiently perform video segmentation with half the computational cost of a comparable, state-of-the-art image segmentation method.
READ FULL TEXT