Spine-Aware Multi-Person 2D and 3D Pose Estimation

Open-source inference library for non-commercial and research use on Linux, Windows and macOS.

Web Demo Code

Note: SpinePose v2.1.0 is now available! This includes 3D pose estimation capabilities from a single camera, along with improved accuracy and robustness. See the release notes for more details.

SpinePose is an inference library for spine-aware 2D human pose estimation in the wild. It provides a simple CLI and Python API for running inference on images and videos. Our models predict the SpineTrack skeleton hierarchy comprising 37 keypoints, including 9 directly along the spine chain in addition to the standard body joints.

For installation instructions and complete API documentation, see the SpinePose GitHub repository.

---

SpinePose can robustly estimate spine-aware full-body 2D poses for multiple people in unconstrained images and videos.

In addition to body and feet keypoints, we track 9 points along the spine chain, which roughly coincide with the centroids of S1, L1, L3, L5, T8, T3, C7, C4, and C1 vertebral bodies¹.

Spine-aware pose estimation provides additional keypoints along the spine chain, enabling better analysis of torso and spine motion. This is particularly beneficial for applications in fitness, sports, and ergonomics.

[1] Inferred spine keypoints represent a learned mapping from whole-body visual cues to a spine-aware skeleton representation, instead of directly corresponding to unobserved anatomical landmarks.

Our models are specifically designed to work in unconstrained settings, including challenging poses and occlusions. They can be used for real-time applications on consumer-grade hardware.

Warning: The default SpinePose models are not intended for clinical use. Use with caution and do not rely on them for critical decisions. For medical applications, please consult with healthcare professionals and use clinically validated tools.

You can either track the whole body or only spine keypoints, with robust performance during dynamic movements.

Our monocular 3D pose estimation model can predict 3D spine-aware poses in either camera coordinates or automatically estimated metric world coordinates, without requiring any additional sensors or multi-view setups.

SpinePose is also robust to lighting changes and challenging scenarios, and works well in low-light conditions and high motion blur, making it suitable for a wide range of real-world applications.

Model Zoo

2D Pose Estimation

Method	Training Data	SpineTrack				SIMSPINE	Usage
		APB	ARB	APS	ARS	AUC
spinepose_v2_small	SpineTrack + SIMSPINE	0.788	0.815	0.920	0.929	0.790	--mode small --model-version v2
spinepose_v2_medium		0.821	0.846	0.928	0.937	0.798	--mode medium --model-version v2
spinepose_v2_large		0.840	0.862	0.917	0.927	0.803	--mode large --model-version v2
spinepose_v1_small	SpineTrack	0.792	0.821	0.896	0.908	0.611	--mode small --model-version v1
spinepose_v1_medium		0.840	0.864	0.914	0.926	0.633	--mode medium --model-version v1
spinepose_v1_large		0.854	0.877	0.910	0.922	0.633	--mode large --model-version v1
spinepose_v1_xlarge		0.759	0.801	0.893	0.910	-	--mode xlarge --model-version v1

Citation

If you use SpinePose in your research, please consider citing the following papers:

• Towards Unconstrained 2D Pose Estimation of the Human Spine
• SIMSPINE: A Biomechanics-Aware Simulation Framework for 3D Spine Motion Annotation and Benchmarking

Preferred citations can be found on the respective paper pages.

Maintained by saifkhichi96 on GitHub.

The website is distributed under different open-source licenses. For more details, see the notice at the bottom of the page.