RESA: Recurrent Feature-Shift Aggregator for Lane Detection

Lane detection is one of the most important tasks in self-driving. Due to various complex scenarios (e.g., severe occlusion, ambiguous lanes, etc.) and sparse supervisory signals inherent lane annotations, task still challenging. Thus, it difficult for ordinary convolutional neural network (CNN) train general scenes catch subtle feature from raw image. In this paper, we present a novel module named REcurrent Feature-Shift Aggregator (RESA) enrich after preliminary extraction with an CNN. RESA takes advantage strong shape priors lanes captures spatial relationships pixels across rows columns. It shifts sliced map recurrently vertical horizontal directions enables each pixel gather global information. can conjecture accurately challenging weak appearance clues by aggregating map. Moreover, propose Bilateral Up-Sampling Decoder that combines coarse-grained fine-detailed features up-sampling stage. recover low-resolution into pixel-wise prediction meticulously. Our method achieves state-of-the-art results on two popular benchmarks (CULane Tusimple). Code has been made available at: https://github.com/ZJULearning/resa.