Laneformer: Object-Aware Row-Column Transformers for Lane Detection

We present Laneformer, a conceptually simple yet powerful transformer-based architecture tailored for lane detection that is long-standing research topic visual perception in autonomous driving. The dominant paradigms rely on purely CNN-based architectures which often fail incorporating relations of long-range points and global contexts induced by surrounding objects (e.g., pedestrians, vehicles). Inspired recent advances the transformer encoder-decoder various vision tasks, we move forwards to design new end-to-end Laneformer revolutionizes conventional transformers into better capturing shape semantic characteristics lanes, with minimal overhead latency. First, coupling deformable pixel-wise self-attention encoder, presents two row column operations efficiently mine point context along shapes. Second, motivated appearing would affect decision predicting segments, further includes detected object instances as extra inputs multi-head attention blocks encoder decoder facilitate sensing contexts. Specifically, bounding box locations are added Key module provide interaction each pixel query while ROI-aligned features inserted Value module. Extensive experiments demonstrate our achieves state-of-the-art performances CULane benchmark, terms 77.1% F1 score. hope effective will serve strong baseline future models detection.