Towards Unconstrained 2D Pose Estimation of the Human Spine
This addresses a crucial need in sports analytics, healthcare, and realistic animation by providing a comprehensive dataset and method for detailed spine articulation, though it is incremental as it builds on existing body pose estimators.
The authors tackled the problem of 2D spine pose estimation in unconstrained settings by introducing SpineTrack, a dataset with 58k annotations, and SpinePose, a method that extends body pose estimators, achieving validated effectiveness for precise estimation in general and sports-specific contexts.
We present SpineTrack, the first comprehensive dataset for 2D spine pose estimation in unconstrained settings, addressing a crucial need in sports analytics, healthcare, and realistic animation. Existing pose datasets often simplify the spine to a single rigid segment, overlooking the nuanced articulation required for accurate motion analysis. In contrast, SpineTrack annotates nine detailed spinal keypoints across two complementary subsets: a synthetic set comprising 25k annotations created using Unreal Engine with biomechanical alignment through OpenSim, and a real-world set comprising over 33k annotations curated via an active learning pipeline that iteratively refines automated annotations with human feedback. This integrated approach ensures anatomically consistent labels at scale, even for challenging, in-the-wild images. We further introduce SpinePose, extending state-of-the-art body pose estimators using knowledge distillation and an anatomical regularization strategy to jointly predict body and spine keypoints. Our experiments in both general and sports-specific contexts validate the effectiveness of SpineTrack for precise spine pose estimation, establishing a robust foundation for future research in advanced biomechanical analysis and 3D spine reconstruction in the wild.