Exploring Compact and Efficient Neural Networks for Real-Time Semantic Segmentation

Real-time semantic segmentation has profound implications in modern automated driving systems. For safe control, automated driving systems usually need accurate and real-time traffic information such as the positions of pedestrians, traffic signs, other cars, etc., which are mainly provided by semantic segmentation neural networks. However, it is challenging to build an efficient network that can achieve the optimal trade-off between inference speed and accuracy, due to the fact that high prediction accuracy neural network models are usually over-parameterized, their sheer sizes prevent them from deploying in resource-constrained systems. On the contrary, small-scale models face the problem of lacking accuracy. To solve this problem, we presented a novel method to build an efficient semantic segmentation network that provides a better balance between inference speed and prediction accuracy than the existing methods. Our model is built upon a twobranch network architecture, which allows our model to capture spatial details and contextual features efficiently. Moreover, we introduce a compact block to further compress and accelerate our model. We also apply the pseudo-labeling method to obtain higher prediction accuracy. Finally, we evaluate our method on public datasets and compare our method with the current state-of-the-art methods. Specifically, our method achieves 79.8% and 78.1% mean intersection over union (mIOU) on Cityscapes validation