RTMO-ORT : RTMO via ONNXRuntime

Lightweight RTMO pose estimation on pure ONNX Runtime.
Tiny Python class + simple CLIs: rtmo-image, rtmo-video, rtmo-webcam.

If this project helps you, please consider starring the repo — it improves discoverability.

---

Install (pick ONE)

1) pip (quickest)

pip install -U pip
# CPU
pip install rtmo-ort[cpu]
# or GPU (CUDA) wheels if available for your platform
pip install rtmo-ort[gpu]

If the pip package isn’t available yet for your platform, use option 2 (clone) or 3 (conda).

2) Clone + editable install (dev mode)

git clone https://github.com/namas191297/rtmo-ort.git
cd rtmo-ort

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -U pip
pip install -e ".[cpu]"   # or ".[gpu]"

3) Conda environment

conda create -n rtmo-ort python=3.9 -y
conda activate rtmo-ort

# Inside this conda env, use pip to install the package
pip install -U pip
pip install rtmo-ort[cpu]   # or rtmo-ort[gpu]

The package does not ship model files. See “Get the models” below.

---

Get the models (required)

You need the RTMO .onnx files on disk. The recommended way is to run the helper script which downloads and places them under models/.

Option A — Recommended script (from repo root)

bash get_models.sh

This creates a layout like:

models/
  rtmo_s_640x640_coco/rtmo_s_640x640_coco.onnx
  rtmo_s_640x640_crowdpose/rtmo_s_640x640_crowdpose.onnx
  rtmo_s_640x640_body7/rtmo_s_640x640_body7.onnx
  rtmo_m_640x640_coco/rtmo_m_640x640_coco.onnx
  rtmo_m_640x640_body7/rtmo_m_640x640_body7.onnx
  rtmo_l_640x640_coco/rtmo_l_640x640_coco.onnx
  rtmo_l_640x640_crowdpose/rtmo_l_640x640_crowdpose.onnx
  rtmo_l_640x640_body7/rtmo_l_640x640_body7.onnx
  rtmo_t_416x416_body7/rtmo_t_416x416_body7.onnx

Option B — Manual download

Download the .onnx files from Releases and place them under models/ with the same folder structure as above.
You can also point to a custom models folder using:

export RTMO_MODELS_DIR=/abs/path/to/my_models

---

Model matrix & direct downloads

Size	Dataset	Input	Keypoints	Download (.onnx)
tiny	body7	416	17	rtmo_t_416x416_body7.onnx
small	coco	640	17	rtmo_s_640x640_coco.onnx
small	crowdpose	640	14	rtmo_s_640x640_crowdpose.onnx
small	body7	640	17	rtmo_s_640x640_body7.onnx
medium	coco	640	17	rtmo_m_640x640_coco.onnx
medium	body7	640	17	rtmo_m_640x640_body7.onnx
large	coco	640	17	rtmo_l_640x640_coco.onnx
large	crowdpose	640	14	rtmo_l_640x640_crowdpose.onnx
large	body7	640	17	rtmo_l_640x640_body7.onnx

The table links always point to the latest Release. If a link 404s, refresh the Release or run ./get_models.sh.

---

Quickstart (CLI)

Image → annotated image

# Small COCO (17 kp), CPU
rtmo-image --model-type small --dataset coco \
  --input assets/demo.jpg --output out.jpg --device cpu

Video → annotated video

# Medium COCO, GPU if available
rtmo-video --model-type medium --dataset coco \
  --input input.mp4 --output out.mp4 --device cuda

Webcam

# Tiny Body7 (coarse skeleton), default camera 0
rtmo-webcam --model-type tiny --dataset body7 --device cpu --cam 0

Useful flags

• --no-letterbox : disable letterboxing (use simple resize)
• --score-thr, --kpt-thr, --max-det : tweak detection & keypoint thresholds
• --size : override input size (defaults: tiny=416, small/medium/large=640)
• --onnx : bypass presets and point directly to a specific .onnx

Preset matrix (what --model-type + --dataset map to):

model-type	dataset	default size	expected K
tiny	body7	416	17
small	coco	640	17
small	crowdpose	640	14
small	body7	640	17
medium	coco/body7	640	17
large	coco/body7	640	17

CrowdPose uses 14 keypoints (we connect top_head + neck + limbs).
COCO uses 17 keypoints.
Body7 exports in this repo contain 17 keypoints (coarser subset represented).

---

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, score_thr=0.25, kpt_thr=0.2)

img = cv2.imread("assets/demo.jpg")
boxes, kpts, scores = pe.infer(img)   # boxes[N,4], kpts[N,K,3], scores[N]

vis = pe.annotate(img, boxes, kpts, scores)
cv2.imwrite("vis.jpg", vis)
print(f"saved vis.jpg — persons: {len(boxes)}")

---

Troubleshooting

• No detections on your image/video? Try lower thresholds:

  --score-thr 0.05 --kpt-thr 0.05

• Webcam issues: pass --cam 1 (or another index), add --window to name the window.
• OpenCV can’t open display on some Linux servers: you may need sudo apt-get install -y libgl1 or run headless and save to disk.
• GPU not used: check printed ONNX Runtime providers. If you only see CPUExecutionProvider, install the GPU wheel (pip install onnxruntime-gpu) and ensure CUDA/cuDNN are available.

---

Notes

• ONNX Runtime version is not pinned; use the latest stable wheel.
• Letterboxing is optional; most RTMO exports assume external resize. --no-letterbox is available to experiment.
• You can store models anywhere and set RTMO_MODELS_DIR to point to that folder.

---

Acknowledgments

• RTMO authors & MMPose contributors.
• ONNX Runtime team for the excellent inference engine.

---

License

Apache-2.0. See LICENSE.