Supplementary Material Rethinking Reprojection: Closing the Loop for Pose-aware Shape Reconstruction from a Single Image

We denote each fully-connected layer fc(d) by its output dimension d, and volumetric convolution layer by conv3D(k, c, s) representing kernel size of k, strides of s across three spatial axes, and c channels. 2D convolutional layer is represented as conv2D(k, c, s). and the volumetric transpose convolution layer by deconv3D(k, c, s). Encoder and Generator for Aligned Shapes The variational aligned shape encoder takes as input an 30×30×30×1 tensor, and consists of 3 convolution layers: conv3D(4, 32, 2), conv3D(4, 64, 2), conv3D(4, 64, 2); two fully-connected layers fc(200) and fc(200), regressing from the last convolution feature to the 200-dimensional mean and variance vectors for style embedding, following [2]. The decoder takes in the 200-dimensional style vector, and consists of one fully-connected layer fc(8192) to connect the input vector to an 4×4×4×128 convolutional feature; and 3 transpose convolution layers deconv3D(4, 64, 2), deconv3D(4, 32, 2), deconv3D(4, 1, 2) output the reconstructed shape with size 30×30×30×1. All convolution and transpose convolution layer are batch batch normalized except the first convolution and last transpose convolution layer. LeakyReLU [3][1] is the rectifier for all layers except the output layer which uses tanh. This architecture is also used for 3D VAE in Section 4.2. Image to Style/Pose Regressors The two regressors have identical architecture of convolution layers: conv2D(11, 64, 4), conv2D(5, 128, 2), conv2D(5, 256, 2), conv2D(5, 512, 2), conv2D(3, 200, 2). For the style regressor, an fc(200) connects the last convolution layer to the style parameters. For the pose regressor, fc(5) is used instead. All but the first convolution layers are batch normalized, and rectified with LeakyReLU.