Labeled Hands in Wild

Hand pose estimation has been playing a major role in many applications such as in Augmented/Virtual reality that is the human-computer interaction and gesture recognition. Among the existing hand datasets, some of them are synthetically generated which do not provide information about the background considering the various lighting conditions where the hand skin tone information would be lost.Hence, the proposed Labelled Hand Dataset in the Wild provides this additional information and also solves a major problem ie occlusion. Since manually annotating a large dataset is a tedious task,hence we propose a novel approach to automate the generation of large dataset using the triangulation method which is also known as multiview Annotation.In this approach two best frames are labelled with the 2D points(21 keypoints) which are then triangulated in 3D space using multiview geometry with the use of fiducial markers.These triangulated points in a 3D space are reprojected onto all other images of a particular pose and this process is repeated for all the other poses thus automating the generation of large labeled dataset in wild.