Guided Dual Network Based Transfer with an Embedded Loss for Lane Detection in Nighttime Scene

摘要

Lane detection in the daytime has achieved great success in recent years. Compared with the daytime, lane detection at nighttime is still more difficult because the light is very weak. The previous methods use the planet of additional annotated data at nighttime environment to train the network. However, the annotation of images at nighttime is not only very expensive, but also very hard to annotate. Consequently, in this paper, we propose a dual network, which takes two images of the same scene in daytime and nighttime as input and detects lanes simultaneously. Specifically, we first adopt the transfer to generate nighttime images from the daytime images in an unsupervised manner. Note that in this way, we do not require annotate any data. Then, we send the daytime and nighttime images of the same scene to a dual network, in which an embedded loss is used to guide nighttime branch to learning discriminative information from the daytime branch to improve the detection of nighttime lanes. To strength the feature extraction of lane detection, we also propose streak-aware pooling to aggregate local contextual information. Extensive experiments on the nighttime scene of the CULane benchmark demonstrates the effectiveness of our method.