Cityscapes TL++: Semantic Traffic Light Annotations for the Cityscapes Dataset

The inconvenient truth of ground truth errors in automotive datasets and DNN-based detection

“…nuScenes has also published instructions for their labelling, however these labels are open to interpretation and produce errors similar to the ones identified into other datsets, e.g. KITTI MoSeg dataset, as shown in Figure 1.Other datasets have been developed and annotated using deep learning methods such as active learning and neural networks [1,18,26]. These methods of annotation are otherwise known as semi-automatic.…”

German Traffic Sign Recognition Using Convolutional Neural Network

“…However, they lack additional attributes such as orientation, pictogram, and relevance information, which are necessary to utilize the detected TLs for autonomous driving. Examples of the datasets extended with TL states include COCO Traffic [12], where TL states were annotated in the images from the COCO [13] dataset, as well as Cityscapes TL++ dataset [10] containing images with fine annotations from the Cityscapes [14] dataset with additional TL labels for four attributes: type (car, pedestrian, bicycle, train, unknown), relevant (yes, no), visible (yes, no), and state (red, red-yellow, yellow, green, off, unknown). Other datasets containing only TL state labels are the Roboflow Self-Driving Car dataset [15], a modified version of the Udacity Self-Driving Car Dataset [16], Waymo Open Dataset [17], WPI [7], BDD100K [18], and ApolloScape [19] datasets.…”