Apppublic

Kawal511/Ergonomics-MultiHMR

sourceHugging Faceotherupdated 22d agoView on Hugging Face

1likes

App README

Ergonomic Analysis

Upload a barbershop photo or clip. Multi-HMR builds a 3D whole-body skeleton of everyone in frame, then a deterministic scoring layer reports:

—Barber (standing) — REBA posture-risk score, Low / Med / High / Critical
—Client (seated) — comfort profile plus chair-design suggestions

No training, no configuration. Roles are assigned automatically.

Deploying this Space

1. Put SMPL-X in a private dataset repo

SMPL-X is licence-restricted and must not be committed to a public Space.

1.<https://huggingface.co/new-dataset> → name it smplx-neutral → visibility Private.
2.Upload your SMPLX_NEUTRAL.npz to it.

2. Create the Space

<https://huggingface.co/new-space> → SDK Gradio → Hardware:

Hardware	Speed	Notes
CPU basic (free)	~20–40 s per photo	works; use `multiHMR_672_S`
ZeroGPU (needs PRO)	~1 s per photo	recommended
T4 small (paid)	~0.5 s per photo	always-on

3. Upload the files

Clone the Space and push app.py, requirements.txt and this README.md:

bash

git clone https://huggingface.co/spaces/YOUR-USERNAME/ergonomic-analysis
cd ergonomic-analysis
# copy the three files in, then:
git add . && git commit -m "initial" && git push

Or use the Files → Add file → Upload files button in the Space's web UI.

4. Set the secret and variables

Space Settings → Variables and secrets:

Name	Type	Value
`HF_TOKEN`	Secret	a read token from <https://huggingface.co/settings/tokens>
`SMPLX_REPO`	Variable	`YOUR-USERNAME/smplx-neutral`
`MODEL_NAME`	Variable	`multiHMR_672_S` (fast) or `multiHMR_896_L` (accurate)

HF_TOKEN must be a Secret, not a Variable — it grants access to your private repo.

5. First build

The first build takes 10–15 minutes: it installs the dependencies, clones Multi-HMR and downloads the checkpoint. Watch the Logs tab. Once you see setup complete, it's live.

Restarts reuse the cached checkpoint and come up in about a minute.

Troubleshooting

Log message	Fix
`HF_TOKEN secret is not set`	Step 4 — and make sure it's a Secret, not a Variable
`401` / `Repository not found`	The token lacks read access to the private dataset repo
`CUDA out of memory`	Set `MODEL_NAME` to `multiHMR_672_S`
Build times out	Normal on the first run — restart the Space, the checkpoint is cached

Licensing

Multi-HMR (NAVER LABS Europe) is licensed for research and non-commercial use only. Commercial deployment requires a separate licence from NAVER. SMPL-X is likewise non-commercial and cannot be redistributed — hence the private-repo arrangement above.

Accuracy

Validated on 3DPW (1355 frames, 24 sequences): median joint-angle error 7.9°, trunk 4.2°, upper arm 5.7°. Neck and elbow carry a larger systematic offset (16–18°) from a known SMPL↔SMPL-X joint-definition difference. REBA tiers are correct 76.5% of the time and always within one tier.