Purpose- Spatial feature is important for scene saliency detection, scene-based visual saliency detection methods fail to incorporate the spatial aspects of 3D scenes. A cube-based method is proposed to improve saliency detection through integrating visual and spatial features in 3D scenes. Design/methodology/approach- In the presented approach, a multi-scale pyramid of cubes is used to organize the image and mesh model of a 3D scene. Each 3D cube in this pyramid represents a space unit similar to a pixel in the multi-scale image pyramid of the image saliency model. In each 3D cube, color, intensity and orientation features are extracted from image and a quantitative concave-convex descriptor is extracted from 3D space. Then Gaussian filter is used to this pyramid of cubes and an extended center-surround difference is introduced to compute the cube-based saliency of the 3D scene. Findings- The precision-recall rate and receiver operating characteristic (ROC) curve is used to evaluate our method and other state-of-art methods. Result shows that our method is better than traditional image-based methods, especially for 3D scene. Originality/value- The paper presents a method which improve image-based visual saliency model