Multi-Scale Recursive Context Aggregation Network for Semantic Segmentation

Convolution operations that are consecutively applied in a typical CNN architecture, cause the loss of original details in input image signals at the cost of extracting new features. Among these details are the coarse patterns the network model tries to capture in deeper layers. However, those coarse details can be easily detected in lower image resolutions and incorporated into the higher level features. Based on this hypothesis, in this study we propose a novel multi-scale multi-input recursive context aggregation network which works on semantic segmentation tasks and show that it outperforms baseline U-Net model by 2% in mIoU on Oxford-IIIT Pet dataset.