Abstract: To address the efficiency and accuracy challenges in 3D mesh saliency detection, this study proposed a skeleton-driven detection method. By leveraging the correspondence between skeleton endpoints, bending points, joint points, and salient regions, candidate regions were identified while non-salient areas were pruned to reduce computational costs. The global representation capability of skeletons was further utilized to conduct multi-scale saliency detection. Finally, the method was integrated with mesh simplification to support large-scale model processing. Experiments on the Schelling and Stanford datasets demonstrated that, compared to existing methods (e.g., mesh saliency, spectral detection, MRSS, and MIMO-GAN), the proposed approach reduced RMSE errors by 23.4%-38.9% and MESH errors by 22.8%-40.3% on average. Notably, it achieved up to 50× speedup when processing the Armadillo model (345 944 faces).