Skip to content
ModelListedExternal
Download

Pozify Exercise Router

This repository contains the Pozify exercise-router artifacts for classifying pose windows as squat, pushup, shoulderpress, or unknown.

Sourcehuggingface.co/NLag/pozify-exercise-router/tree/5d9eeffd2ca05d3f18910f7fdbb1cb2453a15774
KindOtherVersionTemporalBgrade

Model source

Kind
Other
File size
1.13 MB
Version
vTemporal
Updated
Oct 10, 2026
Source
Hugging Face

Source description

Pozify Exercise Router

This repository contains the Pozify exercise-router artifacts for classifying pose windows as squat, push_up, shoulder_press, or unknown.

Model Details

The active artifact is selected by router_selection.json.

Current selected artifact:

{
  "selected_model": "temporal.pt",
  "selected_artifact": "temporal.pt",
  "reason": "prefer BiLSTM temporal when available; baseline falls back when temporal is missing"
}

Artifacts:

  • temporal.pt: selected PyTorch BiLSTM temporal model trained over 30-frame feature tensors.
  • router_selection.json: active artifact selector used by Pozify runtime loading.
  • router.joblib: scikit-learn baseline artifact kept for comparison and fallback.
  • training_report.md: training and evaluation metrics.

Intended Use

The router is intended for Pozify's local app pipeline. It routes normalized pose sequences to the appropriate exercise-specific analyzer or rejects unsupported/uncertain clips as unknown.

Supported labels:

  • squat
  • push_up
  • shoulder_press
  • unknown

Training Data

Primary source:

  • RickyRiccio/Real_Time_Exercise_Recognition_Dataset

Unsupported classes from the source dataset, including curl variations, are mapped to unknown. Custom unknown clips can include idle standing, setup motion, stretching, partial reps, severe occlusion, and bad camera angles.

Features

The router uses 30-frame sliding windows with engineered pose features:

  • normalized landmarks
  • landmark visibility
  • knee, hip, elbow, and shoulder angles
  • relative distances such as hand width over shoulder width
  • frame deltas and velocities

Evaluation

The latest training report is included as training_report.md.

Summary:

ModelArtifactAccuracyUnknown rejection rate
Baselinebaseline.joblib0.99870.9984
BiLSTM temporaltemporal.pt0.99640.9984

Limitations

  • Metrics are based on the current router-window cache, not a broad deployment benchmark.
  • The router expects usable pose extraction and full-body framing where relevant.
  • Unsupported exercises are intentionally routed to unknown.
  • Additional independent held-out videos are needed before treating this as production-grade.

Runtime Loading

Pozify can load this repository by setting:

export POZIFY_ROUTER_HF_REPO_ID=NLag/pozify-exercise-router

For private repositories, authenticate with hf auth login or set HF_TOKEN.

Sources

1 source
Hugging FacePrimaryVersion 5d9eeffd2ca05d3f18910f7fdbb1cb2453a15774
huggingface.co/NLag/pozify-exercise-router/tree/5d9eeffd2ca05d3f18910f7fdbb1cb2453a15774

Verified Oct 10

Model artifacts

1 artifact

temporal.pt

pt · 1.13 MB · SHA-256 8d906d03d110…0e89 · Hugging Face

Download file

Source excerpts

2 excerpts
ReadmeAug 2huggingface.co/NLag/pozify-exercise-router

--- license: other library_name: scikit-learn tags: - pose-estimation - exercise-recognition - video-classification - pozify datasets: - RickyRiccio/Real_Time_Exercise_Recognition_Dataset --- # Pozify Exercise Router This repository contains the Pozify exercise-router artifacts for classifying pose windows as `squat`, `push_up`, `shoulder_press`, or `unknown`. ## Model Details The active artifact is selected by `router_selection.json`. Current selected artifact: ```json { "selected_model": "temporal.pt", "selected_artifact": "temporal.pt", "reason": "prefer BiLSTM temporal when available; baseline falls back when temporal is missing" } ``` Artifacts: - `temporal.pt`: selected PyTorch BiLSTM temporal model trained over 30-frame feature tensors. - `router_selection.json`: active artifact selector used by Pozify runtime loading. - `router.joblib`: scikit-learn baseline artifact kept for comparison and fallback. - `training_report.md`: training and evaluation metrics. ## Intended Use The router is intended for Pozify's local app pipeline. It routes normalized pose sequences to the appropriate exercise-specific analyzer or rejects unsupported/uncertain clips as `unknown`. Supported labels: - `squat` - `push_up` - `shoulder_press` - `unknown` ## Training Data Primary source: - `RickyRiccio/Real_Time_Exercise_Recognition_Dataset` Unsupported classes from the source dataset, including curl variations, are mapped to `unknown`. Custom unknown clips can include idle standing, setup motion, stretching, partial reps, severe occlusion, and bad camera angles. ## Features The router uses 30-frame sliding windows with engineered pose features: - normalized landmarks - landmark visibility - knee, hip, elbow, and shoulder angles - relative distances such as hand width over shoulder width - frame deltas and velocities ## Evaluation The latest training report is included as `training_report.md`. Summary: | Model | Artifact | Accuracy | Unknown rejection rate | | --- | --- | ---: | ---: | | Baseline | `baseline.joblib` | 0.9987 | 0.9984 | | BiLSTM temporal | `temporal.pt` | 0.9964 | 0.9984 | ## Limitations - Metrics are based on the current router-window cache, not a broad deployment benchmark. - The router expects usable pose extraction and full-body framing where relevant. - Unsupported exercises are intentionally routed to `unknown`. - Addit...

NoteAug 2huggingface.co/NLag/pozify-exercise-router

Source context: 0 downloads · 0 likes · Pipeline video-classification · Library scikit-learn · Repo NLag/pozify-exercise-router