Borrowing a Pose Model for a New Job: Squats, Table Tennis, and Fall Risk

Take a general-purpose pose estimator, bolt a task-specific stage on top, and point it at a new problem. Three 2026 studies do exactly that — grading squats, measuring a table-tennis elbow, predicting a fall — and each one hits a similar wall. The pose backbone travels well; the accuracy number travels with a specific camera setup, not with the method.

One model, many borrowed uses

The recipe is now familiar: an off-the-shelf pose backbone (BlazePose, OpenPose and kin) supplies the keypoints, and a task-specific layer on top turns them into whatever you need — a technique score, a joint angle, a risk flag. It is efficient and it works. But each study is validated in one setup, and the moment you move off that axis — a new camera angle, occlusion, a population outside the test set — the reported accuracy stops being a promise.

Title slide reading 'Borrowing a Pose Model for a New Job — Squats, Table Tennis, and Fall Risk'.
Three 2026 studies bolt a task-specific stage onto a general-purpose pose estimator — and hit a similar wall each time.

Paper 01 — Grading a squat from phone video

The first study pulls BlazePose landmarks from single-camera lift videos and scores squat, bench press and deadlift by the error in their joint-angle trajectories against a reference rep. The result is meaningful for squat and deadlift — movements whose key angles live largely in the camera plane — and weaker for the bench press. A generic pose model plus a simple angle-error readout is enough to grade technique, within its filmed viewpoint.

Pipeline diagram for grading a squat from phone video: phone video, then BlazePose landmarks, then joint angles, then compare against a reference rep.
Paper 01 — a generic pose backbone (BlazePose) with a task-specific readout bolted on top: phone video to a technique score.

Paper 02 — A table-tennis elbow, from two cameras

The second reconstructs a 3D elbow angle from two-camera OpenPose, comparing three camera placements and checking agreement against a marker system via dynamic time warping. Similarity ranged from 0.72 to 0.95 — good, but with a wide spread that depended heavily on where the cameras stood. The method is sound; the number is a property of the placement.

Paper 03 — One second before the fall

The third repairs a noisy monocular 3D pose with RePoseNet, then feeds trunk lean and stride variability into StaFallNet, which flags instability about one second before a fall begins. Trained and tested on the fall subset of UCF101 across 11 sport categories, it is a striking demonstration — but, like the others, its numbers describe its dataset, not the open world.

Same wall, different sport

Three sports, three task-specific heads, one shared wall: each pairs a generic pose model with a bespoke readout, and each degrades off-axis, under occlusion, or outside its test set. The accuracy figures describe one camera setup, not the method behind them — which is the number worth remembering when you borrow a pose model for a job it was not trained for. See also our look at pose-estimation accuracy across six models and the purpose-built CalTennis sports-AI benchmark.

Summary slide 'Same Wall, Different Sport': each study pairs a generic pose model with a task-specific readout; all three degrade off-axis, under occlusion, or outside their test set.
The shared wall: accuracy numbers describe one camera setup, not the method.

Frequently asked questions

Can you grade exercise technique from a phone video?

Yes, within limits. A 2026 study used BlazePose landmarks from single-camera lifts and scored squat and deadlift meaningfully by joint-angle-trajectory error. It works best for movements whose key angles sit in the camera plane; accuracy drops for lifts and angles that do not.

How accurate is two-camera markerless 3D for joint angles?

In a table-tennis study, two-camera OpenPose reconstruction agreed with a marker system at a similarity of 0.72–0.95 for elbow angle — good overall, but with a wide spread that depended on camera placement. The takeaway: the setup, not just the software, sets the accuracy.

Can pose estimation predict a fall before it happens?

One 2026 system (RePoseNet + StaFallNet) flagged instability about a second before fall onset from monocular 3D pose, using trunk lean and stride variability. It is promising on its test set (a fall subset of UCF101), but that number reflects the dataset, not guaranteed real-world lead time.

References

[1] Turner, J., Clark, J., & Keshtmand, N. (2026). Markerless Pose Estimation for Resistance Training Technique Assessment. arXiv:2608.24384.
[2] Cao, Z., Zhang, Z., & Xiao, Y. (2026). Evaluating the accuracy of table tennis 3D markerless motion capture based on two video cameras: a pilot study. Frontiers in Bioengineering and Biotechnology.
[3] He, L., Wang, S., Seo, H., et al. (2026). Two-stage posture enhancement and stability modelling for pre-fall risk detection in athletic movements. Frontiers in Bioengineering and Biotechnology.


Takashi Fukushima — Sports Science & Pose Estimation.
Subscribe on YouTube  ·  Website  ·  ORCID  ·  Contact

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top