We present a semi-automatic annotation method to build a large dataset for traffic landmark detection, where traffic landmarks include traffic signs, traffic lights as well as road markings. Labor-intensive bounding box tagging is a huge challenge to generate a large dataset for detection algorithms. To mitigate the labor, we adopt a high-definition (HD) map and a positioning system. We propose a process to align the HD map and images semi-automatically. Through the registration, the annotations of the HD map can be directly tagged onto traffic landmarks in the images. To make full use of the HD map for the dataset generation, we annotate the traffic landmarks with reference points, following the way that they are represented in the HD map, instead of the bounding boxes. The proposed semi-automatic method speeds up the annotation by a factor of 3.19, as compared to the manual annotation. Our dataset consists of about 150,000 images and includes about 470,000 annotated traffic landmarks. We train a deep neural network on our dataset to detect the traffic landmarks, and its performance is evaluated using a novel evaluation metric. Moreover, we show that the pretrained traffic landmark detection network is effective in detecting traffic landmarks in other countries using the bounding box by fine-tuning.

中文翻译：

---

交通地标注释的半自动框架

我们提出一种半自动注释方法，以构建用于交通标志检测的大型数据集，其中交通标志包括交通标志，交通信号灯和道路标记。劳动密集型的边界框标记是为检测算法生成大型数据集的巨大挑战。为了减轻工作量，我们采用了高清（HD）地图和定位系统。我们提出了一种半自动对齐高清地图和图像的过程。通过注册，可以将HD地图的注释直接标记到图像中的路标上。为了充分利用高清地图进行数据集生成，我们将按照参考点在高清地图中（而不是边界框）的显示方式为交通地标添加参考点。所提出的半自动方法将注释速度提高了3倍。19，与手动注释相比。我们的数据集包含大约150,000张图像，并包含大约470,000条带注释的交通地标。我们在数据集上训练了一个深度神经网络来检测交通标志，并使用一种新颖的评估指标对其性能进行评估。此外，我们表明，使用边界框通过微调，预训练的交通地标检测网络可有效检测其他国家的交通地标。