# RTMO-ORT — RTMO pose on pure ONNX Runtime
Minimal, fast RTMO (person detection + 2D pose) inference with **no heavy frameworks**. One tiny Python class, three simple CLIs, and ready-to-download ONNX models.
If this saves you time, please consider **starring the repo** — it really helps.
---
## Install (two ways)
### A) pip (recommended)
```bash
# CPU
pip install "rtmo-ort[cpu]"
# GPU (uses onnxruntime-gpu if present)
pip install "rtmo-ort[gpu]"
```
### B) From source
```bash
git clone https://github.com/namas191297/rtmo-ort.git
cd rtmo-ort
python -m venv .venv && source .venv/bin/activate # optional
pip install -e ".[cpu]" # or ".[gpu]"
```
Python ≥ 3.8. Works on Linux/macOS/Windows.
---
## Get models
This repo ships a helper to fetch ONNX files into `models/…`.
```bash
# fetch a specific release tag (e.g., v0.1.0)
./get_models.sh v0.1.0
# or omit to use the default in the script
./get_models.sh
```
You can also download individual models manually (see table below).
By default, the CLIs look in `models/`. To change that, set `RTMO_MODELS_DIR=/path/to/models`.
---
## Models table (direct downloads)
Each file should be placed at `models/<name>/<name>.onnx`.
Example: `models/rtmo_s_640x640_coco/rtmo_s_640x640_coco.onnx`.
> Replace `v0.1.0` with your chosen tag if needed.
| Size | Dataset | Input | Download |
|:------:|:----------------|:----:|:--|
| tiny | body7 | 416 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_t_416x416_body7.onnx |
| small | coco | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_s_640x640_coco.onnx |
| small | crowdpose | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_s_640x640_crowdpose.onnx |
| small | body7 | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_s_640x640_body7.onnx |
| medium | coco | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_m_640x640_coco.onnx |
| medium | body7 | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_m_640x640_body7.onnx |
| large | coco | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_coco.onnx |
| large | crowdpose | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_crowdpose.onnx |
| large | body7 | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_body7.onnx |
| large | body7_crowdpose | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_body7_crowdpose.onnx |
---
## Use the CLIs
All commands accept the same presets and thresholds:
- `--model-type {tiny,small,medium,large}` (default: `small`)
- `--dataset {coco,crowdpose,body7,body7_crowdpose}` (default: `coco`)
- `--no-letterbox` (disable square letterbox; default is letterbox on)
- `--score-thr`, `--kpt-thr`, `--max-det`
- `--device {cpu,cuda}`
- `--onnx /path/to/model.onnx` (overrides presets)
- `--models-dir /path/to/models` (default: `models`)
### Image
```bash
rtmo-image --model-type small --dataset coco \
--input path/to/in.jpg --output out.jpg --device cpu
```
### Video
```bash
rtmo-video --model-type small --dataset coco \
--input in.mp4 --output out.mp4 --device cuda
```
### Webcam
```bash
rtmo-webcam --model-type small --dataset coco --device cpu
# pick another camera:
# rtmo-webcam --cam 1
```
---
## Python API
```python
import cv2
from rtmo_ort import PoseEstimatorORT
onnx = "models/rtmo_s_640x640_coco/rtmo_s_640x640_coco.onnx"
pe = PoseEstimatorORT(onnx, device="cpu", letterbox=True)
img = cv2.imread("assets/demo.jpg")
boxes, kpts, scores = pe.infer(img)
vis = pe.annotate(img, boxes, kpts, scores)
cv2.imwrite("vis.jpg", vis)
```
**Outputs**
- `boxes`: `[N,4]` in xyxy
- `kpts`: `[N,K,3]` with `(x,y,score)` per keypoint
- `scores`: `[N]` person scores
---
## Notes and tips
- **NMS is fused** inside the ONNX models. Do not run NMS again.
- **Letterbox vs. stretch.** Letterbox (default) preserves aspect ratio and generally matches training; stretching (`--no-letterbox`) may reduce accuracy but can be fine for quick demos.
- **Keypoints count.** COCO = 17; CrowdPose = 14; Body7 is coarse. Some Body7 exports use 17-dim outputs for compatibility; semantics remain coarse.
- **GPU provider.** If you installed `onnxruntime-gpu`, use `--device cuda`. If CUDA isn’t found, ONNX Runtime falls back to CPU.
- **Codecs.** If `rtmo-video` fails to write a file, try `mp4v`, `XVID`, or install OS-level codecs.
---
## Project structure
```
rtmo_ort/
├─ estimator.py # PoseEstimatorORT (ONNX Runtime + postprocess + drawing)
├─ cli.py # rtmo-image / rtmo-video / rtmo-webcam
└─ __init__.py
models/ # place ONNX files here (or use --onnx)
get_models.sh # fetches model files for a given tag
```
---
If you want something specific, open an issue.
---
## Contributing & support
Issues and pull requests are welcome. If you found this useful, **star the repo** and consider sharing a short demo clip — it helps others discover it.
- Website: namasbhandari.in
- Repo: https://github.com/namas191297/rtmo-ort
- Issues: https://github.com/namas191297/rtmo-ort/issues
- License: Apache-2.0
Raw data
{
"_id": null,
"home_page": null,
"name": "rtmo-ort",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.8",
"maintainer_email": null,
"keywords": "RTMO, keypoints, mmpose, onnxruntime, pose, realtime",
"author": "Namas",
"author_email": null,
"download_url": "https://files.pythonhosted.org/packages/43/e9/2eee2e14770d29ae7dadf433709c9c67c20ef75ae1676ed6974fd95db7e0/rtmo_ort-0.1.0.post1.tar.gz",
"platform": null,
"description": "# RTMO-ORT \u2014 RTMO pose on pure ONNX Runtime\n\nMinimal, fast RTMO (person detection + 2D pose) inference with **no heavy frameworks**. One tiny Python class, three simple CLIs, and ready-to-download ONNX models.\n\nIf this saves you time, please consider **starring the repo** \u2014 it really helps.\n\n---\n\n## Install (two ways)\n\n### A) pip (recommended)\n```bash\n# CPU\npip install \"rtmo-ort[cpu]\"\n\n# GPU (uses onnxruntime-gpu if present)\npip install \"rtmo-ort[gpu]\"\n```\n\n### B) From source\n```bash\ngit clone https://github.com/namas191297/rtmo-ort.git\ncd rtmo-ort\npython -m venv .venv && source .venv/bin/activate # optional\npip install -e \".[cpu]\" # or \".[gpu]\"\n```\n\nPython \u2265 3.8. Works on Linux/macOS/Windows.\n\n---\n\n## Get models\n\nThis repo ships a helper to fetch ONNX files into `models/\u2026`.\n\n```bash\n# fetch a specific release tag (e.g., v0.1.0)\n./get_models.sh v0.1.0\n\n# or omit to use the default in the script\n./get_models.sh\n```\n\nYou can also download individual models manually (see table below). \nBy default, the CLIs look in `models/`. To change that, set `RTMO_MODELS_DIR=/path/to/models`.\n\n---\n\n## Models table (direct downloads)\n\nEach file should be placed at `models/<name>/<name>.onnx`. \nExample: `models/rtmo_s_640x640_coco/rtmo_s_640x640_coco.onnx`.\n\n> Replace `v0.1.0` with your chosen tag if needed.\n\n| Size | Dataset | Input | Download |\n|:------:|:----------------|:----:|:--|\n| tiny | body7 | 416 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_t_416x416_body7.onnx |\n| small | coco | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_s_640x640_coco.onnx |\n| small | crowdpose | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_s_640x640_crowdpose.onnx |\n| small | body7 | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_s_640x640_body7.onnx |\n| medium | coco | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_m_640x640_coco.onnx |\n| medium | body7 | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_m_640x640_body7.onnx |\n| large | coco | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_coco.onnx |\n| large | crowdpose | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_crowdpose.onnx |\n| large | body7 | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_body7.onnx |\n| large | body7_crowdpose | 640 | https://github.com/namas191297/rtmo-ort/releases/download/v0.1.0/rtmo_l_640x640_body7_crowdpose.onnx |\n\n---\n\n## Use the CLIs\n\nAll commands accept the same presets and thresholds:\n\n- `--model-type {tiny,small,medium,large}` (default: `small`)\n- `--dataset {coco,crowdpose,body7,body7_crowdpose}` (default: `coco`)\n- `--no-letterbox` (disable square letterbox; default is letterbox on)\n- `--score-thr`, `--kpt-thr`, `--max-det`\n- `--device {cpu,cuda}`\n- `--onnx /path/to/model.onnx` (overrides presets)\n- `--models-dir /path/to/models` (default: `models`)\n\n### Image\n```bash\nrtmo-image --model-type small --dataset coco \\\n --input path/to/in.jpg --output out.jpg --device cpu\n```\n\n### Video\n```bash\nrtmo-video --model-type small --dataset coco \\\n --input in.mp4 --output out.mp4 --device cuda\n```\n\n### Webcam\n```bash\nrtmo-webcam --model-type small --dataset coco --device cpu\n# pick another camera:\n# rtmo-webcam --cam 1\n```\n\n---\n\n## Python API\n\n```python\nimport cv2\nfrom rtmo_ort import PoseEstimatorORT\n\nonnx = \"models/rtmo_s_640x640_coco/rtmo_s_640x640_coco.onnx\"\npe = PoseEstimatorORT(onnx, device=\"cpu\", letterbox=True)\n\nimg = cv2.imread(\"assets/demo.jpg\")\nboxes, kpts, scores = pe.infer(img)\n\nvis = pe.annotate(img, boxes, kpts, scores)\ncv2.imwrite(\"vis.jpg\", vis)\n```\n\n**Outputs**\n- `boxes`: `[N,4]` in xyxy\n- `kpts`: `[N,K,3]` with `(x,y,score)` per keypoint\n- `scores`: `[N]` person scores\n\n---\n\n## Notes and tips\n\n- **NMS is fused** inside the ONNX models. Do not run NMS again.\n- **Letterbox vs. stretch.** Letterbox (default) preserves aspect ratio and generally matches training; stretching (`--no-letterbox`) may reduce accuracy but can be fine for quick demos.\n- **Keypoints count.** COCO = 17; CrowdPose = 14; Body7 is coarse. Some Body7 exports use 17-dim outputs for compatibility; semantics remain coarse.\n- **GPU provider.** If you installed `onnxruntime-gpu`, use `--device cuda`. If CUDA isn\u2019t found, ONNX Runtime falls back to CPU.\n- **Codecs.** If `rtmo-video` fails to write a file, try `mp4v`, `XVID`, or install OS-level codecs.\n\n---\n\n## Project structure\n\n```\nrtmo_ort/\n \u251c\u2500 estimator.py # PoseEstimatorORT (ONNX Runtime + postprocess + drawing)\n \u251c\u2500 cli.py # rtmo-image / rtmo-video / rtmo-webcam\n \u2514\u2500 __init__.py\nmodels/ # place ONNX files here (or use --onnx)\nget_models.sh # fetches model files for a given tag\n```\n\n---\n\nIf you want something specific, open an issue.\n\n---\n\n## Contributing & support\n\nIssues and pull requests are welcome. If you found this useful, **star the repo** and consider sharing a short demo clip \u2014 it helps others discover it.\n\n- Website: namasbhandari.in\n- Repo: https://github.com/namas191297/rtmo-ort \n- Issues: https://github.com/namas191297/rtmo-ort/issues \n- License: Apache-2.0",
"bugtrack_url": null,
"license": null,
"summary": "RTMO pose estimation with pure ONNX Runtime (tiny class + CLI).",
"version": "0.1.0.post1",
"project_urls": {
"Homepage": "https://github.com/namas191297/rtmo-ort",
"Issues": "https://github.com/namas191297/rtmo-ort/issues"
},
"split_keywords": [
"rtmo",
" keypoints",
" mmpose",
" onnxruntime",
" pose",
" realtime"
],
"urls": [
{
"comment_text": null,
"digests": {
"blake2b_256": "fd14551a1922c37b6dcebc3c1ff57c2a3f5f73e2304542aeb5005485beacc6b6",
"md5": "d4a50d89aa77670fdc4ba2d65987bc81",
"sha256": "4fd71980d2762555f5ce159be9193cd1dd82d8e45af1462b8091bc10b795555c"
},
"downloads": -1,
"filename": "rtmo_ort-0.1.0.post1-py3-none-any.whl",
"has_sig": false,
"md5_digest": "d4a50d89aa77670fdc4ba2d65987bc81",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.8",
"size": 11887,
"upload_time": "2025-08-29T19:34:50",
"upload_time_iso_8601": "2025-08-29T19:34:50.300898Z",
"url": "https://files.pythonhosted.org/packages/fd/14/551a1922c37b6dcebc3c1ff57c2a3f5f73e2304542aeb5005485beacc6b6/rtmo_ort-0.1.0.post1-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": null,
"digests": {
"blake2b_256": "43e92eee2e14770d29ae7dadf433709c9c67c20ef75ae1676ed6974fd95db7e0",
"md5": "ff172e2f4fe39964d1c93f05ebe5beaa",
"sha256": "dab94ada57abccc4a898380fcaf3510d2d15759f034c7528db0d2f6b1144fe5e"
},
"downloads": -1,
"filename": "rtmo_ort-0.1.0.post1.tar.gz",
"has_sig": false,
"md5_digest": "ff172e2f4fe39964d1c93f05ebe5beaa",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.8",
"size": 8252,
"upload_time": "2025-08-29T19:34:51",
"upload_time_iso_8601": "2025-08-29T19:34:51.227632Z",
"url": "https://files.pythonhosted.org/packages/43/e9/2eee2e14770d29ae7dadf433709c9c67c20ef75ae1676ed6974fd95db7e0/rtmo_ort-0.1.0.post1.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2025-08-29 19:34:51",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "namas191297",
"github_project": "rtmo-ort",
"travis_ci": false,
"coveralls": false,
"github_actions": false,
"lcname": "rtmo-ort"
}
JSON
