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)

# CPU
pip install "rtmo-ort[cpu]"

# GPU (uses onnxruntime-gpu if present)
pip install "rtmo-ort[gpu]"

B) From source

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/….

# 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

rtmo-image --model-type small --dataset coco \
  --input path/to/in.jpg --output out.jpg --device cpu

Video

rtmo-video --model-type small --dataset coco \
  --input in.mp4 --output out.mp4 --device cuda

Webcam

rtmo-webcam --model-type small --dataset coco --device cpu
# pick another camera:
# rtmo-webcam --cam 1

---

Python API

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