Dynamic Die-Forging Scene Semantic Segmentation via Point Cloud-BEV Feature Fusion with Star Encoding

Semantic segmentation of workpieces and die cavities is critical for intelligent process monitoring and quality control in hammer die-forging. However, the field of 3D point cloud segmentation currently faces prominent limitations in forging scenario adaptation: existing state-of-the-art (SOTA) methods are predominantly optimized for road driving or indoor scenes, where targets have stable poses and regular surfaces. They lack dedicated designs for capturing the fine-grained deformation characteristics of forging workpieces and alleviating multi-scale feature misalignment caused by large pose variations-key pain points in forging segmentation. Consequently, these methods fail to balance segmentation accuracy and real-time efficiency required for practical forging applications. To address this gap, this paper proposes a novel semantic segmentation framework fusing 3D point cloud and bird's-eye-view (BEV) representations for complex die-forging scenes. Specifically, a Star-based encoding module is designed in the BEV encoding stage to enhance capture of fine-grained workpiece deformation characteristics. A hierarchical feature-offset alignment mechanism is developed in decoding to alleviate multi-scale spatial and semantic misalignment, facilitating efficient cross-layer fusion. Additionally, a weighted adaptive fusion module enables complementary information interaction between point cloud and BEV modalities to improve precision.We evaluate the proposed method on our self-constructed simulated and real die-forging point cloud datasets. The results show that when trained solely on simulated data and tested directly in real-world scenarios, our method achieves an mIoU that surpasses RPVNet by 1.1%. After fine-tuning with a small amount of real data, the mIoU further improves by 5%, reaching optimal performance.