Depth-embedded multiple pooling for image classification

Most existing methods of image classification ignore the role of depth information hidden in 2-D images. However, the depth information is important for visual perception, especially when the appearance information does not perform well. In this paper, we propose to embed depth information within multiple pooling into the classic platform of image classification, namely bag-of-features. The proposed method quantifies depth diversity by projecting objects to their nearby depth planes, resulting pooling features in the 3-D space indirectly. Experimental results on the MIT Indoor Scene database demonstrate that our proposed depth-embedded multiple pooling is effective to enhance the accuracy of image classification, especially when the appearance features alone are not so discriminative.