Human Pose Estimation Experiments

Note: This repository is under active development. Results and notebook flows are updated as new benchmark runs are completed.

This repository is a practical benchmark and analysis workspace for MediaPipe, OpenPose, AlphaPose, and Detectron2. It gives one consistent pipeline for cross-framework comparison, canonical keypoint export, and motion-signal analysis with reproducible artifacts.

What this repo focuses on

• Running multiple pose frameworks with one benchmark contract.
• Mapping outputs to one canonical COCO-17 subset.
• Exporting stable CSV/JSON artifacts for downstream tasks.
• Building lightweight motion features from keypoint time series.

Why this repo

Pose frameworks expose different keypoint definitions, confidence semantics, and output formats. Without a canonical mapping and export contract, speed and behavior comparisons are easy to misread.

This repository keeps comparison honest by enforcing shared schema, shared artifact structure, and explicit not_measured status whenever a tool cannot be measured in the current environment.

What you will find here

• Canonical mapping utilities under src/posebench/keypoints_schema.py.
• Frame-level exporters with stable columns in src/posebench/export.py.
• Feature extraction utilities in src/posebench/features.py.
• Benchmark runner with environment capture in scripts/run_benchmarks.py.
• Colab-first notebooks for per-tool demos and cross-tool analysis.

Repository layout

.
├── MediaPipe/
├── OpenPose/
├── AlphaPose/
├── Detectron2/
├── notebooks/
├── src/posebench/
├── scripts/
├── results/
├── assets/
├── docs/
└── tests/

Notebook index

Notebook	Scope	Open in Colab
MediaPipe/01_mediapipe_pose_demo.ipynb	MediaPipe inference, canonical mapping, export, mini benchmark	Open
MediaPipe/02_mediapipe_export_and_features.ipynb	Frame sequence export, angles, smoothing, velocity	Open
Detectron2/01_detectron2_keypoints_demo.ipynb	Detectron2 keypoint demo and export	Open
OpenPose/01_openpose_install_and_run.ipynb	OpenPose recommended path and synthetic fallback	Open
AlphaPose/01_alphapose_colab_inference.ipynb	AlphaPose recommended path and synthetic fallback	Open
notebooks/01_benchmark_all_tools.ipynb	Runs benchmark script and inspects generated artifacts	Open
notebooks/02_keypoints_timeseries_analysis.ipynb	Time-series angles, angular velocity, and compact features	Open
notebooks/03_quality_metrics_without_ground_truth.ipynb	Missing-rate, confidence, and temporal jitter diagnostics	Open

Results snapshot

Generated from results/benchmark.csv. Only measured tools contain numbers.

Tool	Status	Avg ms/frame	Std ms/frame	FPS
mediapipe	measured	7.41	0.16	134.95
detectron2	measured	1033.24	47.56	0.97
openpose	measured	429.79	5.90	2.33
alphapose	not_measured	-	-	-

Full table and notes are in results/benchmark.md.

Quick start

Colab

• Open a notebook from the table above.
• Setup cells clone the repo when needed and install dependencies idempotently.

Local

git clone https://github.com/sumeyye-agac/human-pose-estimation-experiments.git
cd human-pose-estimation-experiments
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements-dev.txt

Validation commands:

ruff check .
pytest -q
python scripts/check_links.py --skip-remote
python scripts/verify_results_consistency.py

Optional framework dependencies

• MediaPipe

pip install mediapipe==0.10.14 "numpy<2"

• Detectron2 (best effort)
  • Colab or Linux GPU flows can use the official install matrix.
  • macOS arm64 CPU can use conda-forge builds.

conda create -y -n posebench-d2 -c conda-forge python=3.10 detectron2
conda run -n posebench-d2 python -m pip install mediapipe==0.10.14 opencv-python-headless "setuptools<81"

• OpenPose and AlphaPose
  • Official installs can be fragile in ephemeral Colab sessions.
  • Notebooks include fallback paths for schema/export validation.
  • Fallback outputs are clearly labeled synthetic and never treated as real inference.

Output format

All tools are mapped into a canonical COCO-17 subset.

• Schema and mapping details are in docs/schema.md.
• CSV contract includes frame_index, timestamp_ms, person_id, tool, schema, and per-keypoint {name}_x, {name}_y, {name}_confidence columns.

Reproducibility and consistency

Benchmark artifacts are generated with:

python scripts/run_benchmarks.py --tool all

Generated artifacts:

• results/benchmark.csv
• results/benchmark.md
• results/environment.json
• results/benchmark_raw_<tool>.json for measured tools

results/environment.json records Python version, platform, optional GPU probe result, and key library versions for the run.

Method details are in docs/benchmark_methodology.md.

Honest results policy

• Measured numbers are published only when a real adapter run finishes in the current environment.
• Unsupported or failed tools stay not_measured with a clear reason.
• Synthetic fallback outputs are labeled as synthetic and used only to validate pipeline contracts.
• No row is backfilled with guessed latency or FPS.

Limitations and next steps

• AlphaPose remains not_measured on macOS arm64 CPU due official CUDA-dependent custom ops.
• OpenPose measurement currently uses OpenCV DNN with official COCO Caffe weights rather than pyopenpose runtime.
• Multi-person identity tracking is not yet part of the shared export contract.
• Current benchmark input is synthetic; next step is fixed real-video clips with published license.
• Cross-machine comparison still needs standardized runtime presets.

References

• MediaPipe Pose
• OpenPose (official)
• AlphaPose (official)
• Detectron2 (official)

Author

• GitHub
• LinkedIn