CANet: Curved Guide Line Network with Adaptive Decoder for Lane Detection

As a crucial and challenging task for automated driving, lane detection has been widely explored from different perspectives, especially in the deep learning era. However, the SOTA approaches are difficult to recognize corner lanes effectively. This paper first proposes ''guide line'' to constrain the lane origins and suggests a U-shaped curved guide line to turn grazing angles bigger for stable learning. By using Gaussian mask in supervision stage, the adaptive decoder mechanism could choose between row- or column-wise classification more intelligently, and the prediction of location and range behave more consistently.