Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation.

Spatial pyramid pooling module or encode-decoder structure are used in deepneural networks for semantic segmentation task. The former networks are able toencode multi-scale contextual information by probing the incoming features withfilters or pooling operations at multiple rates and multiple effectivefields-of-view, while the latter networks can capture sharper object boundariesby gradually recovering the spatial information. In this work, we propose tocombine the advantages from both methods. Specifically, our proposed model,DeepLabv3+, extends DeepLabv3 by adding a simple yet effective decoder moduleto refine the segmentation results especially along object boundaries. Wefurther explore the Xception model and apply the depthwise separableconvolution to both Atrous Spatial Pyramid Pooling and decoder modules,resulting in a faster and stronger encoder-decoder network. We demonstrate theeffectiveness of the proposed model on the PASCAL VOC 2012 semantic imagesegmentation dataset and achieve a performance of 89% on the test set withoutany post-processing. Our paper is accompanied with a publicly availablereference implementation of the proposed models in Tensorflow.