Deep Learning Features Inspired Saliency Detection of 3D Images

Abstract

Saliency detection of 3D images is important for many 3D applications, such as bit allocation in 3D video coding, spatial pooling in stereoscopic image quality assessment and feature extraction in 3D object retrieval. However, traditional saliency detection approaches only target for the 2D images. Meanwhile, the traditional hand-crafted low-level feature extraction process may be not suitable for the 3D images. In this paper, we propose a deep learning feature based 3D visual saliency detection model. The pre-trained CNN model is employed to extract the feature vectors for both color and depth images after multi-level image segmentation. Then, we train a neutral network based classifier to generate the color and depth saliency maps from the feature vectors. Final, the linear fusion method is adopted to obtain the final saliency map for 3D image. Experimental results demonstrate that our proposed model can achieve appealing performance improvement over two public benchmark datasets.