You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Yes — ruvnet/wifi-densepose-mmfi-pose is published by this project, under the same Hugging Face account that owns this repository.
It is the 17-keypoint pose model documented in the repository:
docs/benchmarks/mmfi-wifi-sensing-study.md:42 — "Published: ruvnet/wifi-densepose-mmfi-pose", the model behind the 82.69 % torso-PCK@20 in-domain figure on the MM-Fi random_split protocol (ensemble 83.59 %).
docs/adr/ADR-187-archive-v1-deprecation-honest-labeling.md:129 — listed under the "Real & validated" tier, alongside ruvnet/wifi-densepose-pretrained.
docs/benchmarks/wifi-pose-efficiency-frontier.md — the efficiency comparison that uses it as the flagship checkpoint.
One boundary worth stating, because the number is easy to misread: 82.69 % is the public-benchmark figure for that published checkpoint. It is not the accuracy of the committed on-device model in v2/crates/cog-pose-estimation, which reaches PCK@20 3.0 % and whose runtime path is still a confidence=0 stub. The README states this distinction in "Model weights: what's real, what's not".
Activity
danalec commentedon Oct 22, 2026
Yes —
ruvnet/wifi-densepose-mmfi-poseis published by this project, under the same Hugging Face account that owns this repository.It is the 17-keypoint pose model documented in the repository:
docs/benchmarks/mmfi-wifi-sensing-study.md:42— "Published:ruvnet/wifi-densepose-mmfi-pose", the model behind the 82.69 % torso-PCK@20 in-domain figure on the MM-Firandom_splitprotocol (ensemble 83.59 %).docs/adr/ADR-187-archive-v1-deprecation-honest-labeling.md:129— listed under the "Real & validated" tier, alongsideruvnet/wifi-densepose-pretrained.docs/benchmarks/wifi-pose-efficiency-frontier.md— the efficiency comparison that uses it as the flagship checkpoint.One boundary worth stating, because the number is easy to misread: 82.69 % is the public-benchmark figure for that published checkpoint. It is not the accuracy of the committed on-device model in
v2/crates/cog-pose-estimation, which reaches PCK@20 3.0 % and whose runtime path is still aconfidence=0stub. The README states this distinction in "Model weights: what's real, what's not".