Abstract: To address the challenges of incomplete edge extraction and high residual noise in existing point cloud object edge detection methods, we propose MESAM-Net—a novel edge detection framework based on Multi-branch Edge Convolution and a Spatial Attention Mechanism. Our approach leverages a multi-branch edge convolution module to extract rich local and global geometric features, which are then effectively integrated via a three-stage feature fusion strategy. Furthermore, a spatial attention mechanism assigns adaptive weights to points based on their contextual importance, enhancing the precision of edge localization. We evaluate MESAM-Net on the ABC dataset, achieving an Accuracy of 97.2% and an IoU of 76.2%. Extensive comparisons and ablation studies further demonstrate the effectiveness and robustness of our method.