| --- |
| license: apache-2.0 |
| library_name: onnx |
| tags: |
| - onnx |
| - pose-estimation |
| - dwpose |
| - rtmpose |
| - rtmw |
| - yolox |
| - mmpose |
| - coreml |
| pipeline_tag: keypoint-detection |
| --- |
| |
| # DWPose (RTMW + YOLOX) — ONNX |
|
|
| The two ONNX graphs behind the **DWPose** whole-body pose detector, mirrored for |
| [SceneWorks](https://github.com/SceneWorks/SceneWorks) so the `pose_detect` job and the |
| OpenPose control-conditioning lanes can be installed from the Model Manager instead of |
| fetching archives mid-job. |
|
|
| This is rtmlib's *performance* preset: a YOLOX-m person detector followed by an RTMW-x |
| COCO-WholeBody-133 SimCC pose model. Both run under `onnxruntime` — CoreML EP on Apple |
| Silicon, CUDA EP (CPU fallback) elsewhere. |
|
|
| | file | role | input | size (bytes) | |
| |---|---|---|---| |
| | `yolox_m_8xb8-300e_humanart-c2c7a14a.onnx` | person boxes (NMS baked in) | `[1,3,640,640]` f32, letterboxed, pad 114 | 101,400,344 | |
| | `rtmw-dw-x-l_simcc-cocktail14_270e-384x288_20231122.onnx` | whole-body 133-keypoint SimCC | `[1,3,384,288]` f32 BGR, mean/std normalized | 229,320,930 | |
|
|
| Both graphs are ONNX IR version 6, `producer: pytorch 1.9`. |
|
|
| ## Provenance |
|
|
| These files are **extracted verbatim, with no modification of any kind**, from the ONNX SDK |
| archives published by OpenMMLab. No re-export, no conversion, no quantization, no graph |
| surgery — the bytes are the `end2end.onnx` entry of each archive, renamed to the upstream |
| archive's stem so the two graphs are distinguishable side by side. |
|
|
| Source archives: |
|
|
| | archive | size (bytes) | sha256 | |
| |---|---|---| |
| | [`yolox_m_8xb8-300e_humanart-c2c7a14a.zip`](https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/onnx_sdk/yolox_m_8xb8-300e_humanart-c2c7a14a.zip) | 94,223,081 | `a000224fd8ba283202bc62d4a5fcdfe353adb9f468777dbac1ea2ada2093adde` | |
| | [`rtmw-dw-x-l_simcc-cocktail14_270e-384x288_20231122.zip`](https://download.openmmlab.com/mmpose/v1/projects/rtmw/onnx_sdk/rtmw-dw-x-l_simcc-cocktail14_270e-384x288_20231122.zip) | 213,433,855 | `a87e1af41a0a067776dba7d46e1c21c8f6e9f18e247e0e606718dd1f31e96ffd` | |
|
|
| Extraction map: |
|
|
| ``` |
| yolox_m_8xb8-300e_humanart-c2c7a14a.zip |
| 20230928/yolox_onnx/yolox_m_8xb8-300e_humanart-c2c7a14a/end2end.onnx |
| -> yolox_m_8xb8-300e_humanart-c2c7a14a.onnx |
| |
| rtmw-dw-x-l_simcc-cocktail14_270e-384x288_20231122.zip |
| end2end.onnx |
| -> rtmw-dw-x-l_simcc-cocktail14_270e-384x288_20231122.onnx |
| ``` |
|
|
| Files in this repo: |
|
|
| ``` |
| sha256 yolox_m_8xb8-300e_humanart-c2c7a14a.onnx 3dea6513388889f0fff4b77bf7a26013600321b9eb9ceb0e9a400a82572f5f23 |
| sha256 rtmw-dw-x-l_simcc-cocktail14_270e-384x288_20231122.onnx bd033156e5104c4f5d2edfe0453e02661e30a2f3da453ec93c8764d561b83054 |
| ``` |
|
|
| The archives dropped alongside each `end2end.onnx` (`deploy.json`, `pipeline.json`, |
| `detail.json`, and two sample renders) are MMDeploy SDK metadata that SceneWorks does not |
| read; they are not mirrored here. Re-download either archive, verify its sha256 above, and |
| take the `end2end.onnx` entry to reproduce these files bit for bit. |
|
|
| Upstream: |
|
|
| - Project: [open-mmlab/mmpose](https://github.com/open-mmlab/mmpose) |
| - RTMW: [`projects/rtmpose`](https://github.com/open-mmlab/mmpose/tree/main/projects/rtmpose) — RTMW whole-body models, `cocktail14` training mixture |
| - Model zoo / download index: [RTMPose model zoo](https://github.com/open-mmlab/mmpose/blob/main/projects/rtmpose/README.md) |
| - DWPose (the preset these two graphs form): [IDEA-Research/DWPose](https://github.com/IDEA-Research/DWPose), as packaged by [rtmlib](https://github.com/Tau-J/rtmlib) |
|
|
| ## License |
|
|
| **Apache-2.0**, following mmpose, whose `LICENSE` (reproduced here verbatim, including its |
| `Copyright 2018-2020 Open-MMLab. All rights reserved.` notice) is the license OpenMMLab |
| publishes these releases under. |
|
|
| Note, without overclaiming: mmpose's `LICENSE` covers the OpenMMLab project. It does not |
| itself assert terms for the training data behind these particular checkpoints — the YOLOX |
| detector is trained on **HumanArt**, and the RTMW pose model on the **COCO-WholeBody |
| "cocktail14"** mixture. Those datasets carry their own terms from their own publishers, |
| which mmpose's LICENSE file does not speak to and which this mirror therefore does not |
| represent. If your use is sensitive to dataset provenance, review the HumanArt and |
| COCO-WholeBody terms directly. |
|
|
