MC-hands-1M: A glove-wearing hand dataset for pose estimation

We introduce MC-hands-1M, a synthetic glove-wearing hand dataset for pose estimation. In the zip folder, there exist two subfolders: one containing roughly 750K images (Big set) and another with 250K images (Small set) along with the 2D camera plane and 3D world ground truth data of the corresponding poses. Each set is organized in folders named as Rendered View X, representing a specific camera in the 3D space with a fixed rotation and location. In each of those folders, there exist a json file containing corresponding data for the camera (location, rotation, intrinsics' matrix and images' relative paths) along with the aforementioned ground truth per image (pose). For each of those views, there exist other subfolders named as Scene 's Collection 's Objects' States' Combination Y. Each of those folders contains images of the different poses from the set camera view, given a different combination of background, lighting, glove- and cloth-like materials, and hand's a priori scaling state.

!!!!! IMPORTANT NOTE !!!!!

After some testing, an error in the ground truth of the small set of images was found. Please use only the big set found in the zip. For further instructions and questions, please contact us at prod@iti.gr.