Supplemental Material: Detailed, accurate, human shape estimation from clothed 3D scan sequences

On BUFF dataset, we use λskin = 100, λoutside = 100, λfit = 3, λcpl = 1, and λprior = 0.1. At the stage of estimating per-frame shape TEst, we increase λskin by a factor 10 to retrieve more personalized details. On INRIA dataset, since texture information is not available, we consider all vertices as cloth and therefore set λskin = 0. We decreased λfit = 1 to be more robust to wide clothing, and keep other weights unchanged. To make a fair comparison in this dataset, we initialize the pose using the exact same Stitched Puppet [4] landmarks computed in [3]. To create the fusion scan from the single-frame objective results, we only retain one frame every 1/3 of second. We found this sampling to be a good trade-off between accuracy and computational time. 1.3. Computation time