README.md

---
license: mit
library_name: acaua
pipeline_tag: keypoint-detection
tags:
  - pose-estimation
  - keypoint-detection
  - vision
  - acaua
  - native-pytorch-port
datasets:
  - coco
---

# UniFormer-S COCO top-down pose — acaua mirror (pure-PyTorch port)

Pure-PyTorch port of **UniFormer-S** (top-down COCO 17-keypoint pose at
256x192) hosted under `CondadosAI/` for use with the
[acaua](https://github.com/CondadosAI/acaua) computer vision library.

The architecture has been re-implemented in pure PyTorch under
`acaua.adapters.uniformer.pose` — no `mmcv`, no `mmengine`, no
`mmpose`, no `trust_remote_code`, no `timm` runtime dependency. The
backbone reuses `UniFormer2DDense` (already shipped via PR #9 for the
Stage 1.5 dense-prediction work); the pose head is a fresh port of
mmpose's `TopDownSimpleHead`. Decode is the upstream
`post_process='default'` path (argmax + 0.25-pixel shift, NOT the DARK
unbiased decoder). Inverse-warp uses the shared
`acaua.pose.topdown_utils` module (introduced in PR #8 ahead of this
stage).

## Provenance

| | |
|---|---|
| Upstream code | [`Sense-X/UniFormer`](https://github.com/Sense-X/UniFormer) @ `main` (Apache-2.0); files derived: `pose_estimation/mmpose/models/backbones/uniformer.py` (backbone, identical to detection variant up to module-class identity) + `pose_estimation/mmpose/models/keypoint_heads/top_down_simple_head.py` (head) |
| Upstream weights | Google Drive file id `162R0JuTpf3gpLe1IK6oxRoQK7JSj4ylx`, filename `top_down_256x192_global_small.pth` (101MB) |
| Upstream SHA256 | `d77059e3e9322c0e20dc89dc0cf2a583ffe2ced7d3e9b350233738add570bc30` |
| Upstream report | AP 74.0 / AP@50 90.3 / AP@75 82.2 on COCO val 2017, 256x192, single-scale |
| Architecture | UniFormer-S backbone (`hybrid=False, windows=False`, depth=[3,4,8,3], embed_dims=[64,128,320,512], head_dim=64) + TopDownSimpleHead (3x ConvTranspose2d-stride-2 + BN+ReLU upsample, 1x1 conv to 17 channels) |
| Total params | 25.23M (backbone 21.04M + head 4.19M) |
| Mirrored on | 2026-04-25 |
| Mirrored by | [CondadosAI/acaua](https://github.com/CondadosAI/acaua) |

## Usage via acaua

```python
import acaua

# MIT-declared weights -> explicit opt-in (same posture as RTMPose +
# UniFormer image / video classifications). The bundled RTMDet-tiny
# detector is loaded automatically from CondadosAI/rtmdet_t_coco.
model = acaua.Model.from_pretrained(
    "CondadosAI/uniformer_s_coco_pose", allow_non_apache=True
)

result = model.predict("image.jpg")
print(result.keypoints.shape)        # (N_persons, 17, 2)
print(result.keypoint_scores.shape)  # (N_persons, 17)

# COCO skeleton edges are surfaced on the adapter:
import supervision as sv
sv.EdgeAnnotator(edges=model.skeleton).annotate(scene, result.to_supervision())
```

## Files in this mirror

- `model.safetensors` — full pose model weights (backbone + head, 352
  tensors). Loaded under `load_state_dict(strict=True)` at adapter
  init time.
- `config.json` — `acaua_task=pose`, COCO-17 `keypoint_names` +
  `skeleton`, `detector_repo_id=CondadosAI/rtmdet_t_coco`. Adapter
  surfaces these as `model.keypoint_names` / `model.skeleton`.
- `NOTICE` — attribution chain (code AND weights).
- `LICENSE` — Apache-2.0.

## License and attribution

The adapter code is redistributed under Apache-2.0. The underlying
weights carry upstream's MIT declaration (compatible). The acaua
UniFormer-pose adapter is a derivative work of the upstream PyTorch
implementation — see [`NOTICE`](./NOTICE) for the attribution chain.

## Citation

```bibtex
@inproceedings{li2022uniformer,
  title     = {UniFormer: Unifying Convolution and Self-attention for Visual Recognition},
  author    = {Li, Kunchang and Wang, Yali and Zhang, Junhao and Gao, Peng and Song, Guanglu and Liu, Yu and Li, Hongsheng and Qiao, Yu},
  booktitle = {ICLR},
  year      = {2022},
}
```