Real-time semantic segmentation is of significant importance for mobile and
robotics related applications. We propose a computationally efficient
segmentation network which we term as ShuffleSeg. The proposed architecture is
based on grouped convolution and channel shuffling in its encoder for improving
the performance. An ablation study of different decoding methods is compared
including Skip architecture, UNet, and Dilation Frontend. Interesting insights
on the speed and accuracy tradeoff is discussed. It is shown that skip
architecture in the decoding method provides the best compromise for the goal
of real-time performance, while it provides adequate accuracy by utilizing
higher resolution feature maps for a more accurate segmentation. ShuffleSeg is
evaluated on CityScapes and compared against the state of the art real-time
segmentation networks. It achieves 2x GFLOPs reduction, while it provides on
par mean intersection over union of 58.3% on CityScapes test set. ShuffleSeg
runs at 15.7 frames per second on NVIDIA Jetson TX2, which makes it of great
potential for real-time applications.

我们提出了 ShuffleSeg，这是一种基于分组卷积和通道混洗的有效分割网络，它采用 Skip 架构作为解码方法，在速度和准确性之间达到了最佳的折中，通过利用分辨率更高的特征图进行更准确的分割，它在 CityScapes 测试集上的平均交集联合达到了 58.3％，并且在 NVIDIA Jetson TX2 上的运行帧率为 15.7，非常适合实时应用。

ShuffleSeg：实时语义分割网络

ShuffleSeg: Real-time Semantic Segmentation Network

Convolutional networks are powerful visual models that yield hierarchies of
features. We show that convolutional networks by themselves, trained
end-to-end, pixels-to-pixels, improve on the previous best result in semantic
segmentation. Our key insight is to build "fully convolutional" networks that
take input of arbitrary size and produce correspondingly-sized output with
efficient inference and learning. We define and detail the space of fully
convolutional networks, explain their application to spatially dense prediction
tasks, and draw connections to prior models. We adapt contemporary
classification networks (AlexNet, the VGG net, and GoogLeNet) into fully
convolutional networks and transfer their learned representations by
fine-tuning to the segmentation task. We then define a skip architecture that
combines semantic information from a deep, coarse layer with appearance
information from a shallow, fine layer to produce accurate and detailed
segmentations. Our fully convolutional network achieves improved segmentation
of PASCAL VOC (30% relative improvement to 67.2% mean IU on 2012), NYUDv2, SIFT
Flow, and PASCAL-Context, while inference takes one tenth of a second for a
typical image.

本文提出了一种基于卷积神经网络的 Fully Convolutional Networks（FCN）方法，可以用于图像的深度语义分割。通过该方法，我们成功地提高了 PASCAL VOC、NYUDv2、SIFT Flow 和 PASCAL-Context 等公开数据集上的 segmentation 模型的性能，并且推断速度远快于之前的相关工作。