Efficient Attention Pyramid Network for Semantic Segmentation

Semantic segmentation is a task that covers most of the perception needs intelligent vehicles in an unified way. Recent studies witnessed attention mechanisms achieve impressive performance computer vision task. Current based methods differ with each other position and form mechanism, perform differently practice. This paper firstly introduces effectiveness multi-scale context features tasks. We find channel can play vital role constructing effective features. Based on this analysis, proposes efficient pyramid network (EAPNet) for semantic segmentation. Specifically, to handle problem segmenting objects at multiple scales, we design (ECAP) which employ atrous convolution cascade or parallel capture by using rates. Furthermore, propose residual fusion block (RAFB), whose purpose simultaneously focus meaningful low-level feature maps spatial location information. At same time, will explore different modules modules, describe their impact performance. empirically evaluate our EAPNet two datasets, including PASCAL VOC 2012 Cityscapes datasets. Experimental results show without MS COCO pre-training any post-processing, achieved 81.7% mIoU validation set. With deeplabv3+ as benchmark, improve model more than 1.50% mIoU.