Supplementary Material Multiview Human Body Reconstruction from Uncalibrated Cameras

We randomly select samples from them with a ratio of 0.8: 0.2 Each method reconstruct a unified mesh from multiview inputs. It comes with ground truth 3D pose and follow its standard split for training and testing. MannequinChallenge is captured by a moving hand-held camera while subjects stay still in daily life scenarios. We use annotations provided by Leroy et al. [ There are 45.3M learnable parameters in our architecture, among them 12.6M are used for the We test run time for a different number of views within a frame. The graph convolutional mesh decoder has 7 layers, each with 256 channels.