GASP: A New Feature Fusion Module Based on ASPP

Abstract

By obtaining multiple-scale context information from the same feature maps, the feature aggregation modules can better express the semantic information, which are beneficial for improving the performance of semantic segmentation models. In addition, compared with small receptive fields, large receptive fields can obtain more comprehensive context information, improving the segmentation accuracy. The ASPP uses global pooling and atrous convolution, and the large receptive field can be obtained by global pooling. However, using the global pooling results in the reduction of feature resolution and features. We proposed the GASP feature aggregation module and demonstrated its performance. The ablation experiments on Cityscapes validation dataset showed that mIoU of the GASP module increased by 0.65% compared with the original ASPP module.