Simple very deep convolutional network for robust hand pose regression from a single depth image

Abstract We propose a novel approach for articulated hand pose estimation from a single depth image using a very deep convolutional network. For the first, a very deep network structure is designed to directly maps a single depth image to its corresponding 3D hand joint locations. This approach eliminates the necessity of hand-crafted intermediate features and sophisticated post-processing stages for robust and accurate hand pose estimation. We use Batch Normalization to accelerate training and prevent the objective function from getting stuck in poor local minima. We introduce a low-dimensional embedding forcing the network to learn the inherent constraints of hand joints, which helps to reduce the cost of reconstructing 3D hand poses from high-dimension feature space. We discuss the effect of the convolutional network depth on its accuracy under the hand pose regression setting. Quantitative assessments on two challenging datasets show that our proposed method gets competitive results to state-of-the-art approaches in terms of accuracy. Moreover, qualitative results also show that our proposed method is robust to some difficult hand poses.