Computer Vision

YOLO Pose Estimation for Fitness Images

A Python training and inference project creates person boxes and body-keypoint overlays on exercise images for visual review and fitness-product development.

Exercise still transformed into a person box and body-keypoint overlay by YOLO pose estimation.

Illustrative still-image inference. YOLO pose estimation produces a person box and body-keypoint overlay for visual inspection. Form scoring remains separate.

Exercise-image pose overlay for visual review. The printed person score is prediction confidence, not a project-wide accuracy result.Exercise-image pose overlay for visual review. The printed person score is prediction confidence, not a project-wide accuracy result.Kneeling-pose example for visual review. Some lower-body keypoints are misplaced despite the printed detection confidence.Exercise-image pose overlay for visual review. The printed person score is prediction confidence, not a project-wide accuracy result.Exercise-image pose overlay for visual review. The printed person score is prediction confidence, not a project-wide accuracy result.Additional supplied exercise-pose overlay in the same style. Its original project file was not recovered; the printed score is prediction confidence.

The problem

A fitness product needed a repeatable way to detect a person and inspect body keypoints in exercise photos. The vision component had to accept a labeled dataset, support pose-model training and save overlays that a developer could review.

What I built

A packaged Python project for training an Ultralytics YOLO pose model on a YOLO-format dataset.
Configurable dataset and pose-checkpoint inputs.
Still-image inference that saves person boxes and body-keypoint overlays.
Sample exercise overlays and commands for repeating training and inference.

The diagram shows image input, pose inference and a reviewable overlay. The gallery includes varied poses so joint placement can be checked visually. Printed detection confidence belongs to an individual prediction; it is not a project-wide accuracy score.

The outcome

The handoff provides a repeatable vision component for generating exercise-pose overlays. Review can reveal misplaced joints before the output is used in a fitness product. The consumer app and automatic exercise-form scoring require additional work.

Where this fits

Useful for fitness and movement-analysis teams developing an image-based pose feature. Share permitted sample images, target poses and the intended use of the keypoints to discuss Computer Vision Model Development.

Working on something similar?

Discuss a similar project