Efficient semantic segmentation with pyramidal fusion

Emergence of large datasets and resilience of convolutional models have enabled successful training of very large semantic segmentation models. However, high capacity implies high computational complexity and therefore hinders real-time operation. We therefore study compact architectures which aim at high accuracy in spite of modest capacity. We propose a novel semantic segmentation approach based on shared pyramidal representation and fusion of heterogeneous features along the upsampling path. The proposed pyramidal fusion approach is especially effective for dense inference in images with large scale variance due to strong regularization effects induced by feature sharing across the resolution pyramid. Interpretation of the decision process suggests that our approach succeeds by acting as a large ensemble of relatively simple models, as well as due to large receptive range and strong gradient flow towards early layers. Our best model achieves 76.4% mIoU on Cityscapes test and runs in real time on low-power embedded devices. (c) 2020 Elsevier Ltd. All rights reserved.