Curvature-Aware Attention Enhanced End-to-End Contour Network

In order to solve the key problems of insufficient representation of complex boundaries and difficult registration of fixed vertices in case segmentation, this paper proposes an end-to-end contour network with enhanced curvature perception attention. Through the two-stage attention architecture, the fine-grained initial contour is generated adaptively by combining the channel and spatial attention mechanism, and the long-range dependence of vertices is modeled by graph-convolution network (GCN), which significantly improves the boundary modeling ability. At the same time, the multi-directional dynamic alignment mechanism is introduced, and the vertex matching between occlusion and deformed scene is adaptively optimized by combining curvature consistency constraints. In addition, the network integrates the relocation optimization strategy, and when the boundary prediction fails, it expands the feature response range through curvature perception search to achieve robust contour restoration. Experiments on KITTI dataset show that the average segmentation accuracy of this algorithm is improved by 3.3% and the average recall rate is improved by 4.5% compared with mainstream methods, and it runs in real time on NVIDIA 3060Ti at 6 Frames Per Second(FPS), which is significantly better than the traditional mask model.