README.md

---
license: apache-2.0
library_name: onnx
pipeline_tag: keypoint-detection
tags:
  - pose-estimation
  - rtmw
  - rtmpose3d
  - whole-body
  - onnx
---

# RTMW3D-X, float16-weight ONNX (for ErgoCapture)

Smaller download of **RTMW3D-X** (`rtmw3d-x_8xb64_cocktail14-384x288`), a 133-keypoint whole-body 3D pose model from
[OpenMMLab MMPose](https://github.com/open-mmlab/mmpose/tree/main/projects/rtmpose3d) (Apache-2.0), based on the ONNX
export at [Soykaf/RTMW3D-x](https://huggingface.co/Soykaf/RTMW3D-x).

Used by [ErgoCapture](https://ergocapture.vercel.app), an in-browser ergonomic posture assessment app.

## What changed

- Weights are **stored** as float16 and cast back to float32 inside the graph; ONNX Runtime folds the casts when the
  session loads, so **computation stays float32**. Download size: 369 MB → 187 MB.
- The SimCC output head (last layers, ~3.7 MB) is kept in float32.
- Inputs and outputs are unchanged (`input` 1×3×384×288 float32; SimCC x / y / z outputs).

## Check against the original

On real image crops, 95.5–100% of SimCC argmax positions are identical and the rest differ by one bin (0.5 px), except
three of 1,197 keypoint axes where the original model itself had two near-equal peaks (a thumb tip with score 0.42, one
toe depth and one face point).

sha256: `dc276b94339c17d459902baad32652606ade51df1f8b3effb96b130b4a9e5f7a`

## License

Apache-2.0, as the original RTMW3D model. Copyright of the original weights: OpenMMLab contributors.