QNet: A Quick Deep Neural Network for Real-Time Semantic Segmentation

Real-time semantic segmentation plays an important role in intelligent industrial scenarios like autonomous driving, and the core challenge is to consider both performance and efficiency. In this paper, we present a quick network called QNet to address this problem. The overall design of the proposed network is an asymmetric encoder-decoder architecture. The encoder is based on a new Bottleneck architecture that combines channel split and channel shuffle, which reduces network parameters while improving the level of extracting features. An improved spatial pyramid multi-feature fusion module is adopted to the decoder, which can enlarge the receptive field and make up for the loss of spatial information, with only a small computational complexity. Experiments on Camvid and Cityscapes datasets show that without using any pre-trained model, the proposed QNet achieves a state-of-the-art trade-off between performance and efficiency, better than other methods. Specifically, in a single GTX 1080Ti GPU, it can run at over 38 FPS on the Camvid test dataset with 51.8% mean IoU, and 37 FPS with 75.28% category IoU on the Cityscapes test dataset.