DriPE: A Dataset for Human Pose Estimation in Real-World Driving Settings
The task of 2D human pose estimation has known a significant gain of performance with the advent of deep learning. This task aims to estimate the body keypoints of people in an image or a video. However, real-life applications of such methods bring new challenges that are under-represented in the general context datasets. For instance, driver status monitoring on consumer road vehicles introduces new difficulties, like self- and background body-part occlusions, varying illumination conditions, cramped view angles, etc. These monitoring conditions are currently absent in general purposes datasets. This paper proposes two main contributions. Firstly, we introduce DriPE (Driver Pose Estimation), a new dataset to foster the development and evaluation of methods for human pose estimation of drivers in consumer vehicles. This is the first publicly available dataset depicting drivers in real scenes. It contains 10k images of 19 different driver subjects, manually annotated with human body keypoints and an object bounding box. Secondly, we propose a new keypoint-based metric for human pose estimation. This metric highlights the limitations of current metrics for HPE evaluation and of current deep neural networks on pose estimation, both on general and driving-related datasets.