PDBNet: Parallel Dual Branch Network for Real-time Semantic Segmentation

To make a trade-off between accuracy and inference speed in real-time applications on the unmanned mobile platform, a novel neural network, named Parallel Dual Branch Network (PDBNet), is proposed. Firstly, a multi-scale module, namely Parallel Dual Branch (PDB), is designed to extract complete information. PDB module consists of two parallel branches to remove detailed low-level information and high-level semantic information while maintaining few parameters. Then, based on the PDB module, PDBNet, a small-scale and shallow structure, is designed for semantic segmentation. A multi-scale module tends to extract abundant information and segment the object out from the image well. The small-scale and shallow structure tends to accelerate the inference speed. So PDBNet architecture is designed to be effective both in terms of accuracy and inference speed. PDBNet adopts three downsamplings to obtain feature maps with high spatial resolution and uses PDB modules with different dilation rates to extract multi-scale features and enlarge the receptive field in the last several layers. Finally, experiments on Camvid dataset and Cityscapes dataset, we respectively get 67.7% and 69.5% Mean Intersection over Union (MIoU) with only 1.82 million parameters and quicker speed on a single GTX 1070Ti card.