This paper introduces Spatial CNN (SCNN), a novel deep learning architecture designed to enhance spatial understanding in traffic scene analysis. SCNN generalizes traditional layer-by-layer convolutions to slice-by-slice convolutions within feature maps, enabling message passing between pixels across rows and columns in a layer. This approach is particularly effective for long continuous shape structures or large objects with strong spatial relationships but limited appearance cues, such as traffic lanes, poles, and walls. SCNN is applied to a newly released challenging traffic lane detection dataset and the Cityscapes dataset. The results show that SCNN significantly improves performance in lane detection, outperforming RNN-based ReNet and MRF+CNN (MRFNet) by 8.7% and 4.6% respectively. SCNN also achieved first place in the TuSimple Benchmark Lane Detection Challenge with an accuracy of 96.53%. The paper discusses the challenges of traffic scene understanding, particularly in detecting objects with strong structure priors but weak appearance cues, such as lane markings that may be occluded. Traditional methods like MRF/CRF and RNN-based approaches are computationally expensive and less effective in such scenarios. SCNN addresses these issues by enabling efficient message passing and residual learning, making it more computationally efficient and effective for spatial relationships. SCNN is evaluated on two tasks: lane detection and semantic segmentation. In lane detection, SCNN outperforms existing methods, including deep residual networks, in preserving the continuity of long thin structures. In semantic segmentation, SCNN improves performance for large objects requiring global context, such as walls, trucks, and buses. The paper also presents an ablation study showing that SCNN's effectiveness is due to its sequential message passing scheme, which reduces redundancy and improves performance. SCNN is shown to be more efficient than LSTM-based methods and significantly faster than dense CRF in terms of computational efficiency. Finally, SCNN is applied to the Cityscapes dataset for semantic segmentation, demonstrating its effectiveness in improving segmentation results for various categories, including long-shaped objects and large objects. The results show that SCNN is a versatile and effective approach for spatial understanding in traffic scenes.This paper introduces Spatial CNN (SCNN), a novel deep learning architecture designed to enhance spatial understanding in traffic scene analysis. SCNN generalizes traditional layer-by-layer convolutions to slice-by-slice convolutions within feature maps, enabling message passing between pixels across rows and columns in a layer. This approach is particularly effective for long continuous shape structures or large objects with strong spatial relationships but limited appearance cues, such as traffic lanes, poles, and walls. SCNN is applied to a newly released challenging traffic lane detection dataset and the Cityscapes dataset. The results show that SCNN significantly improves performance in lane detection, outperforming RNN-based ReNet and MRF+CNN (MRFNet) by 8.7% and 4.6% respectively. SCNN also achieved first place in the TuSimple Benchmark Lane Detection Challenge with an accuracy of 96.53%. The paper discusses the challenges of traffic scene understanding, particularly in detecting objects with strong structure priors but weak appearance cues, such as lane markings that may be occluded. Traditional methods like MRF/CRF and RNN-based approaches are computationally expensive and less effective in such scenarios. SCNN addresses these issues by enabling efficient message passing and residual learning, making it more computationally efficient and effective for spatial relationships. SCNN is evaluated on two tasks: lane detection and semantic segmentation. In lane detection, SCNN outperforms existing methods, including deep residual networks, in preserving the continuity of long thin structures. In semantic segmentation, SCNN improves performance for large objects requiring global context, such as walls, trucks, and buses. The paper also presents an ablation study showing that SCNN's effectiveness is due to its sequential message passing scheme, which reduces redundancy and improves performance. SCNN is shown to be more efficient than LSTM-based methods and significantly faster than dense CRF in terms of computational efficiency. Finally, SCNN is applied to the Cityscapes dataset for semantic segmentation, demonstrating its effectiveness in improving segmentation results for various categories, including long-shaped objects and large objects. The results show that SCNN is a versatile and effective approach for spatial understanding in traffic scenes.