Region ensemble network: Towards good practices for deep 3D hand pose estimation

3D hand pose estimation is an important and challenging problem for human-computer interaction. Recently convolutional networks (ConvNet) with sophisticated design have been employed to address it, but the improvement is not so significant. To exploit good practice and promote the performance for hand pose estimation, we propose a Region Ensemble Network (REN) for directly 3D coordinate regression. It first partitions the last convolutional outputs of ConvNet into several grid regions. Results from separate fully-connected (FC) regressors on each regions are integrated by another FC layer to perform estimation. By exploitation of several training strategies including data augmentation and smooth L1 loss, REN significantly improves the performance of ConvNet for hand pose estimation. Experiments demonstrate that our approach achieves strong performance on par or better than state-of-the-art algorithms on three public hand pose datasets. We also experiment our methods on fingertip detection and human pose datasets and obtain state-of-the-art accuracy.