We present a comprehensive framework for egocentric interaction recognition\nusing markerless 3D annotations of two hands manipulating objects. To this end,\nwe propose a method to create a unified dataset for egocentric 3D interaction\nrecognition. Our method produces annotations of the 3D pose of two hands and\nthe 6D pose of the manipulated objects, along with their interaction labels for\neach frame. Our dataset, called H2O (2 Hands and Objects), provides\nsynchronized multi-view RGB-D images, interaction labels, object classes,\nground-truth 3D poses for left & right hands, 6D object poses, ground-truth\ncamera poses, object meshes and scene point clouds. To the best of our\nknowledge, this is the first benchmark that enables the study of first-person\nactions with the use of the pose of both left and right hands manipulating\nobjects and presents an unprecedented level of detail for egocentric 3D\ninteraction recognition. We further propose the method to predict interaction\nclasses by estimating the 3D pose of two hands and the 6D pose of the\nmanipulated objects, jointly from RGB images. Our method models both inter- and\nintra-dependencies between both hands and objects by learning the topology of a\ngraph convolutional network that predicts interactions. We show that our method\nfacilitated by this dataset establishes a strong baseline for joint hand-object\npose estimation and achieves state-of-the-art accuracy for first person\ninteraction recognition.\n