Feasibility test

One photo in. Any camera angle out.

Testing whether we can turn ordinary footage of a real teacher into owned 3D pose assets — renderable from any angle, with an anatomy layer video can't show. Everything below was produced end-to-end on this machine with free, open-source tools. No commercial mocap service was used.

01The pipeline

Five stages. The only human input is one photograph.

Step 1

Source

One photo of a real teacher. CC-licensed here; our own footage in production.

Step 2

Detect

MediaPipe extracts 33 3D body landmarks from the single image.

Step 3

Retarget

Landmarks mapped onto a fixed-proportion skeleton — keep the angles, correct the limb lengths.

Step 4

Build

A 3D figure is generated in Blender from the corrected skeleton.

Step 5

Render

Any camera angle, plus an alignment/anatomy layer.

02The input

A single photograph of Warrior II, and the 33 landmarks detected from it. All 33 came back above the confidence threshold — the tracking is accurate against the real body.

Source photograph of Warrior II — Source photographVirabhadrasana II · Lilian White · CC BY 4.0 via Wikimedia Commons

Detected pose landmarks overlaid on the photograph — Detected skeleton, overlaid33/33 landmarks above confidence threshold

03Four angles from that one photo

The photographer stood in one place. These four viewpoints did not exist in the source image — they are rendered from the reconstructed 3D pose. A student can walk around the pose; a video can never offer that.

04The alignment layer

The same pose with the spine line drawn and the joints a teacher actually checks — knees, hips, shoulders — highlighted. This is the teaching advantage: showing why a pose works, not just what it looks like.

Anatomy layer, front — FrontSpine line + checkpoints

Anatomy layer, three-quarter — Three-quarter

05Honest read

What this proves

The pipeline works end-to-end, with free tools, on commodity hardware
Landmark detection on a clear standing pose is accurate
Novel camera angles are genuinely reconstructed, not faked
The anatomy layer is real and is the strongest teaching differentiator
Marginal cost per additional pose is close to zero once set up
Assets are owned outright and reusable across certification and all 187 pose pages

What it does not prove yet

The rear foot floats. There is no inverse-kinematics ground contact, so retargeted limbs don't reliably land on the floor
The figure is primitive geometry — a real rigged mesh with muscle layers is a separate step
Single-image depth is approximate; multi-frame video will be materially better
Floor poses and heavy occlusion will need manual cleanup, as expected
Every pose still requires faculty sign-off — 3D can be wrong as easily as AI imagery

Recommendation: the concept is validated — this is worth building. The remaining gaps (ground contact, a proper rigged mesh) are exactly what commercial retargeting tools solve for roughly $20–100/month, rather than problems requiring original research. Next step before committing: run the same test on video of a teacher rather than a still, and evaluate one paid tool against this free baseline.