Supplemental Material: Hierarchical Kinematic Human Mesh Recovery

– Encoder: We use the standard ResNet50, giving a 2048-dimensional feature vector. – Chains Qc: Each Qc is implemented with a set of fully connected layers. The ψ embedding module Ec comprises two fully connected units (with ReLU activations and dropout in training) with 32-dimensional outputs each. The input dimensionality of Ec varies according to the chain. This is 2070 for the root chain, 2085 for the arm chains, 2082 for the leg chains, and 2076 for the head chain. Finally, each ∆θi is realized with one single fully connected layer with a 3-dimensional output. – Number of inner iterations: 4 (so 2 forward-backward cycles). – Number of outer iterations T = 3. – Shape estimating neural network: Three fully connected units (with ReLU activations and dropout during training) with output units of 512, 128, and 10 respectively. – Camera estimating neural network: Three fully connected units (with ReLU activations and dropout during training) with output units of 512, 128, and 3 respectively. – VAE encoder: Input dimensionality of 69 (23 joint pose parameters); two fully connected units (with Leaky ReLU activation, batch normalization, and dropout in training) with output units of 512 and 32 each, so the latent space dimensionality is 32. – VAE decoder: Input dimensionality of 32 (latent space vector); two fully connected units (with Leaky ReLU activation, batch normalization, and dropout in training) with output units of 512 and 69 each, so the output dimensionality is 69. – HKMR training parameters. We set the loss weights λsmpl = 1, λ2D = 1, λ3D = 1, and λKL = 0.001. We use a batch size of 128, a learning rate of 3e − 4 and the Adam optimizer in training. In each batch, half of the