Akada Hiroyasu ; Wang Jian ; Shimada Soshi ; Takahashi Masaki ; Theobalt Christian ; Golyanik Vladislav

Abstract

We present UnrealEgo, i.e., a new large-scale naturalistic dataset foregocentric 3D human pose estimation. UnrealEgo is based on an advanced conceptof eyeglasses equipped with two fisheye cameras that can be used inunconstrained environments. We design their virtual prototype and attach themto 3D human models for stereo view capture. We next generate a large corpus ofhuman motions. As a consequence, UnrealEgo is the first dataset to providein-the-wild stereo images with the largest variety of motions among existingegocentric datasets. Furthermore, we propose a new benchmark method with asimple but effective idea of devising a 2D keypoint estimation module forstereo inputs to improve 3D human pose estimation. The extensive experimentsshow that our approach outperforms the previous state-of-the-art methodsqualitatively and quantitatively. UnrealEgo and our source codes are availableon our project web page.