The paper introduces Spatial CNN (SCNN), a novel approach that generalizes traditional deep convolutional neural networks (CNNs) to slice-by-slice convolutions within feature maps. This enables the propagation of information between pixels across rows and columns, enhancing the ability to capture spatial relationships in images. SCNN is particularly effective for detecting long continuous structures or large objects with strong spatial relationships but weak appearance clues, such as traffic lanes, poles, and walls. The authors evaluate SCNN on a challenging traffic lane detection dataset and the Cityscapes dataset, demonstrating significant improvements over existing methods. SCNN outperforms recurrent neural networks (RNNs) and Markov Random Fields (MRFs) in lane detection tasks, achieving an accuracy of 96.53% on the TuSimple Benchmark Lane Detection Challenge. The paper also highlights the computational efficiency and flexibility of SCNN, making it suitable for integration into various layers of deep neural networks.The paper introduces Spatial CNN (SCNN), a novel approach that generalizes traditional deep convolutional neural networks (CNNs) to slice-by-slice convolutions within feature maps. This enables the propagation of information between pixels across rows and columns, enhancing the ability to capture spatial relationships in images. SCNN is particularly effective for detecting long continuous structures or large objects with strong spatial relationships but weak appearance clues, such as traffic lanes, poles, and walls. The authors evaluate SCNN on a challenging traffic lane detection dataset and the Cityscapes dataset, demonstrating significant improvements over existing methods. SCNN outperforms recurrent neural networks (RNNs) and Markov Random Fields (MRFs) in lane detection tasks, achieving an accuracy of 96.53% on the TuSimple Benchmark Lane Detection Challenge. The paper also highlights the computational efficiency and flexibility of SCNN, making it suitable for integration into various layers of deep neural networks.