Learning Hierarchical Feature Representation In Depth Image

This paper presents a novel descriptor, geodesic invariant feature (GIF), for representing objects in depth images. Especially in the context of parts classification of articulated objects, it is capable of encoding the invariance of local structures effectively and efficiently. The contributions of this paper lie in our multi-level feature extraction hierarchy. (1) Low-level feature encodes the invariance to articulation. Geodesic gradient is introduced, which is covariant with the non-rigid deformation of objects and is utilized to rectify the feature extraction process. (2) Mid-level feature reduces the noise and improves the efficiency. With unsupervised clustering, the primitives of objects are changed from pixels to superpixels. The benefit is two-fold: firstly, superpixel reduces the effect of the noise introduced by depth sensors; secondly, the processing speed can be improved by a big margin. (3) High-level feature captures nonlinear dependencies between the dimensions. Deep network is utilized to discover the high-level feature representation. As the feature propagates towards the deeper layers of the network, the ability of the feature capturing the data's underlying regularities is improved. Comparisons with the state-of-the-art methods reveal the superiority of the proposed method.