Mucha Wiktor ; Wray Michael ; Kampel Martin

Abstract

Hand pose represents key information for action recognition in the egocentricperspective, where the user is interacting with objects. We propose to improveegocentric 3D hand pose estimation based on RGB frames only by usingpseudo-depth images. Incorporating state-of-the-art single RGB image depthestimation techniques, we generate pseudo-depth representations of the framesand use distance knowledge to segment irrelevant parts of the scene. Theresulting depth maps are then used as segmentation masks for the RGB frames.Experimental results on H2O Dataset confirm the high accuracy of the estimatedpose with our method in an action recognition task. The 3D hand pose, togetherwith information from object detection, is processed by a transformer-basedaction recognition network, resulting in an accuracy of 91.73%, outperformingall state-of-the-art methods. Estimations of 3D hand pose result in competitiveperformance with existing methods with a mean pose error of 28.66 mm. Thismethod opens up new possibilities for employing distance information inegocentric 3D hand pose estimation without relying on depth sensors.