ABSTRACT

Real-time semantic segmentation aims to generate high-quality prediction in limited time. Recently, with the development of many related potential applications, such as autonomous driving, robot sensing and augmented reality devices, semantic segmentation is desirable to make a trade-off between accuracy and inference speed with limited computation resources. This paper introduces a novel effective and light-weighted network based on Separable Pyramid Module (SPM) to achieve competitive accuracy and inference speed with fewer parameters and computation. Our proposed SPM unit utilises factorised convolution and dilated convolution in the form of a feature pyramid to build a bottleneck structure, which extracts local and context information in a simple but effective way. Experiments on Cityscapes and Camvid datasets demonstrate our superior trade-off between speed and precision. Without pre-training or any additional processing, our SPMNet achieves 71.22% mIoU on Cityscapes test set at the speed of 94 FPS on a single GTX 1080Ti GPU card.

摘要

实时语义分割旨在在有限的时间内生成高质量的预测。最近，随着许多相关潜在应用的发展，例如自动驾驶、机器人传感和增强现实设备，语义分割是在计算资源有限的情况下在准确性和推理速度之间进行权衡的理想选择。本文介绍了一种基于可分离金字塔模块（SPM）的新型高效轻量级网络，以更少的参数和计算量实现具有竞争力的准确性和推理速度。我们提出的 SPM 单元利用特征金字塔形式的分解卷积和扩张卷积来构建瓶颈结构，以简单但有效的方式提取局部和上下文信息。Cityscapes 和 Camvid 数据集的实验证明了我们在速度和精度之间的出色权衡。在没有预训练或任何额外处理的情况下，我们的 SPMNet 在单张 GTX 1080Ti GPU 卡上以 94 FPS 的速度在 Cityscapes 测试集上实现了 71.22% mIoU。