SelfPose: 3D Egocentric Pose Estimation From a Headset Mounted Camera

Abstract

We present a new solution to egocentric 3D body pose estimation from monocular images captured from a downward looking fish-eye camera installed on the rim of a head mounted virtual reality device. This unusual viewpoint leads to images with unique visual appearance, characterized by severe self-occlusions and strong perspective distortions that result in a drastic difference in resolution between lower and upper body. We propose a new encoder-decoder architecture with a novel multi-branch decoder designed specifically to account for the varying uncertainty in 2D joint locations. Our quantitative evaluation, both on synthetic and real-world datasets, shows that our strategy leads to substantial improvements in accuracy over state of the art egocentric pose estimation approaches. To tackle the severe lack of labelled training data for egocentric 3D pose estimation we also introduced a large-scale photo-realistic synthetic dataset. xR-EgoPose offers 383K frames of high quality renderings of people with diverse skin tones, body shapes and clothing, in a variety of backgrounds and lighting conditions, performing a range of actions. Our experiments show that the high variability in our new synthetic training corpus leads to good generalization to real world footage and to state of the art results on real world datasets with ground truth. Moreover, an evaluation on the Human3.6M benchmark shows that the performance of our method is on par with top performing approaches on the more classic problem of 3D human pose from a third person viewpoint.

Silhouette-Based Synthetic Data Generation For 3D Human Pose Estimation With A Single Wrist-Mounted 360° Camera

“…Previous researchers used multiple body-mounted cameras for whole-body pose estimation [9] and used RGB-D cameras for upper-body (i.e., hands, arms, torso) motion estimation [10]. In recent years, methods for whole-body 3D pose estimation with more practical settings have been proposed, such as using a wearable camera with a wide viewing angle to capture more body parts [1][2][3][4]11]. Moreover, several studies have achieved 3D pose estimation under the severe condition in which the human body is completely hidden from the camera's line of sight [5,6,12].…”

“…The majority of recent wearable camera-based 3D human pose estimation methods used fisheye cameras, expecting more body parts to be captured [1][2][3][4]. As the conventional datasets for 3D human pose estimation cannot be directly applied to these fisheye camera-based methods, these methods use synthetic datasets dedicated to each method to train their network.…”

“…Especially, 3D human pose estimation with wearable cameras is key to many important applications, such as lifelogging in terms of medical assistance, monitoring for life support, virtual reality, and sports activity analysis. Several methods that use wearable cameras mounted on the head or chest have been proposed recently [1][2][3][4][5][6]. However, thus far, there is no method with a more practical camera setting, i.e., a single wrist-mounted camera which could be introduced in smartwatches in the future.…”

“…The difficulty is how to prepare the training data; there is no existing dataset for 3D human pose estimation with wrist-mounted cameras. Existing works with body-mounted cameras that faced this issue tackled it by using synthetic data [1][2][3][4]. However, as widely known in these previous works, there is a domain gap issue between real-world data and synthetic data.…”

EgoFish3D: Egocentric 3D Pose Estimation from a Fisheye Camera via Self-Supervised Learning

“…Furthermore, there is a lack of real-world datasets with accurate ground truth data as the collection using a motion capture (MoCap) system is labor-intensive and limited to small laboratory settings. To partially address these issues, recent effort within the vision community has been directed to building public datasets using synthetic human models [13] [14]. However, training with synthetic datasets can affect the generalization capability of the model when subsequently applied to real-world scenarios.…”