3D Convolutional Neural Networks for Efficient and Robust Hand Pose Estimation from Single Depth Images Supplementary Material

We present all the 3D CNN models in the experiment section. We experiment with projective D-TSDF volumes with different resolution values: 16, 32 and 64. Figure 1a presents the network architecture when the input is projective D-TSDF volumes with 32×32×32 resolution. We use this 3D CNN model to compare with state-of-the-art methods on the MSRA dataset [2] and the NYU dataset [3]. However, when the volume resolution is 16×16×16 or 64×64×64, the network architecture is different with that in Figure 1a. Figure 1b presents the network architecture when the input is projective D-TSDF volumes with 16×16×16 resolution. We reduce a convolutional layer and a max pooling layer in this network. Figure 1c presents the network architecture when the input is projective D-TSDF volumes with 64×64×64 resolution. We add a convolutional layer and a max pooling layer in this network. We also experiment with different TSDF types: accurate TSDF, projective TSDF and projective D-TSDF. When the input volume is accurate/projective TSDF which has only one channel, the parameters of the network architecture in Figure 1a should be modified to adapt to the input with one channel. Figure 1d presents the network architecture when the input is accurate/projective TSDF volumes with 32×32×32 resolution. Since the number of input channel is 1 instead of 3, we divide the numbers of output channels for the convolutional layers by 3.