Hand-3D-Studio: A New Multi-View System For 3D Hand Reconstruction

This paper proposes a new system named as Hand-3D-Studio to capture the 3D hand pose and shape information. Our system includes 15 synchronized DSLR cameras, which can acquire high quality multi-view 4K resolution color images in a circular manner. We then introduce a 2D hand keypoints guided iterative pixel growth matching strategy for 3D reconstruction, where the 2D keypoints are obtained via convolution neural network. We find that the pre-detected 2D hand keypoints can greatly remove the matching noise, and thus improve the performance of reconstruction. After that, a non-rigid iterative closest points algorithm is performed to drive a template hand to fit the point clouds and register all the hand meshes. As a consequence, we captured more than 20K high quality hand color images, annotated 2D hand keypoints, 3D point cloud as well as the registered hand meshes (>200). All the data will be public on the website http://www.yangangwang.com for future research.