Understanding the effective receptive field in semantic image segmentation

Abstract

Deep convolutional neural networks trained with strong pixel-level supervision have recently significantly boosted the performance in semantic image segmentation. The receptive field is a crucial issue in such visual tasks, as the output must capture enough information about large objects to make a better decision. In DCNNs, the theoretical receptive field size could be very large, but the effective receptive field may be quite small. The latter is an really important factor in performance. In this work, we defined a method of measuring effective receptive field. We observed that stacking layers with large receptive field can increase the size of receptive field and increase the density of receptive field. Based on the observation, we designed a Dense Global Context Module, which makes the effective receptive field coverage larger and density higher. With the Dense Global Context Module, segmentation model reduces a large number of parameters while the performance has been substantially improved. Massive experiments proved that our Dense Global Context Module exhibits very excellent performance on the PASCAL VOC2012 and PASCAL CONTEXT data set.