EgoGlass: Egocentric-View Human Pose Estimation From an Eyeglass Frame

UnrealEgo: A New Dataset for Robust Egocentric 3D Human Motion Capture

“…5). In extensive experiments, we observe that this simple but effective architecture brings significant improvements compared with existing methods [63,72] qualitatively and quantitatively by 13.5% on MPJPE and 14.65% on PA-MPJPE metrics.…”

“…Egocentric 3D human pose estimation has been actively researched recently [59,69,42,63,62,71,68,72]. Compared to cumbersome motion capture systems that require a fixed recording volume, the egocentric setup is more suitable to capture daily human activities in unconstrained environments.…”

“…Although these methods show promising results, their setups are still not satisfactory for daily use; the cameras are mounted far from the user's body, which is inconvenient and restrictive. The recently introduced EgoGlass approach [72] tackles this issue by an eyeglasses-based setup with two cameras attached to the glasses frame. Their setup imposes fewer restrictions on users' activities.…”

“…Furthermore, existing egocentric datasets are limited in several ways and cannot be easily re-purposed for 3D human pose estimation with the compact eyeglasses-based setup. First, the existing datasets do not contain complex human motions (such as breakdance and backflip) that are seen in daily human activities [59,69,63,72]. Second, the available egocentric datasets do not faithfully model the 3D environment [69,63].…”

“…Second, the available egocentric datasets do not faithfully model the 3D environment [69,63]. Next, the existing stereobased datasets [59,72] do not contain in-the-wild images. All in all, we note that there is no large-scale stereo-based dataset currently available.…”