PoseAtlas / README.md
wjmenu's picture
Update venue to HOUR Workshop @ WACV 2027
43b6c03 verified
|
Raw History Blame Contribute Delete
6.04 kB
---
license: other
license_name: poseatlas-research-use
pretty_name: PoseAtlas
task_categories:
- other
tags:
- synthetic-data
- 3d-pose-estimation
- human-pose
- smpl-x
- computer-vision
- multi-view
size_categories:
- 1M<n<10M
---
# Dataset Card for PoseAtlas
> **Status:** Accepted at the [HOUR Workshop @ WACV 2027](https://hour-workshop.github.io/wacv2027/) (Dataset Track). Dataset files are coming soon and will be
> uploaded to this repository.
>
> [Project page](https://wjmenu.github.io/PoseAtlas/) · [GitHub](https://github.com/wjmenu/PoseAtlas)
## Dataset Summary
PoseAtlas is a large-scale, multi-view synthetic human dataset for 2D and 3D pose estimation.
It provides calibrated multi-view imagery with pose and camera annotations derived directly from
3D geometry rather than inferred post-hoc. Each identity is synthesized in a canonical T-pose,
posed with a target pose sampled from the AMASS motion-capture archive, instantiated as a 3D
avatar, rendered from multiple calibrated viewpoints, and refined by a dual spatial-control
flow-matching model (conditioned on projected avatar geometry and body-part segmentation) trained
with a GRPO-style pose/preference reward.
- **Total images:** 1,000,120
- **Synthesized identities:** 47,625
- **Calibrated viewpoints per pose:** 8
- **Image resolution:** 1024 × 2048
- **Annotations:** SMPL-X body parameters + camera calibration, derived from the underlying 3D geometry
## Supported Tasks
- 2D keypoint / whole-body pose estimation
- 3D human pose and shape estimation (e.g. SMPLer-X–style training protocols)
- Multi-view human reconstruction and avatar modeling
- Synthetic-to-real domain gap / generative data quality research
## Dataset Structure
Each sample corresponds to one `(identity, pose, viewpoint)` triple:
```text
D = { (x_ikj, y_ikj) | i = 1..n identities, k = 1..m poses per identity, j = 1..l views per pose }
```
- `x_ikj`: RGB image (H×W×3) for identity *i*, pose *k*, view *j*
- `y_ikj`: corresponding ground-truth SMPL-X parameters and camera calibration
Planned release fields (subject to change before final release):
| Field | Description |
|---|---|
| `image` | Rendered RGB image, 1024×2048 |
| `identity_id` | Synthesized identity index |
| `pose_id` | AMASS-sourced target pose index |
| `view_id` | Calibrated viewpoint index (of 8) |
| `smplx_params` | SMPL-X pose/shape/expression parameters |
| `camera` | Camera intrinsics/extrinsics for the view |
| `pncc_map` / `segmentation` | Spatial control maps used during generation (PNCC, body-part segmentation) |
## Dataset Creation
### Source Data
- **Identities:** synthesized from attribute prompts (age, ethnicity, gender presentation,
clothing, hairstyle, body shape) using a FLUX-based flow-matching generator conditioned on a
PNCC map of a canonical T-pose SMPL-X body. No real human photographs, scans, or biometric data
are used.
- **Poses:** sampled from the AMASS motion-capture archive and retargeted onto each identity's
SMPL-X shape.
- **Multi-view rendering:** a feed-forward avatar reconstruction model (LHM++) instantiates each
posed identity as 3D Gaussians anchored to the canonical SMPL-X surface, rendered from 8
calibrated orbital viewpoints.
- **Refinement:** a dual spatial-control flow-matching transformer refines each raw render,
conditioned on PNCC geometry, body-part segmentation, and the attribute prompt, trained with a
GRPO objective combining a pose-consistency reward (keypoints vs. geometric GT) and a human
preference reward.
### Annotations
All annotations (SMPL-X parameters, camera calibration) are derived directly from the 3D avatar
geometry used to render each view, rather than estimated from the rendered pixels — the same
geometric reference is shared across all 8 views of a given pose.
## Evaluation Summary (from the accompanying paper)
- **3D pose transfer (SMPLer-X protocol):** a PoseAtlas + BEDLAM training mixture (70/30)
performs comparably to the strongest training mixtures on AGORA and 3DPW (within 0.3 mm of
the lowest error), reaching 77.3 PVE / 107.1 MPE on EHF vs. 78.7 / 108.3 for the closest
alternative.
- **2D pose fidelity:** independent 2D keypoint estimators achieve mean OKS of 0.528–0.798
against geometry-projected keypoints (Sapiens-2B highest across all regions).
- **Multi-view 3D consistency:** MPJPE reaches 24.2 mm using all 8 calibrated views, improving
monotonically as more views are added.
- **Visual quality:** Inception Score 9.58, FID 1.82, comparable to the strongest evaluated
synthetic human dataset baseline.
## Considerations for Using the Data
### Limitations
- Fine-grained **facial identity** varies slightly across views (body geometry, pose, and
clothing remain consistent, since they share explicit SMPL-X geometry; facial appearance is
reconstructed independently per view).
- **Single-person only** — the dataset does not currently model multi-person interaction,
overlap, or occlusion.
### Ethical Considerations
PoseAtlas uses entirely synthetic human identities and does not incorporate photographs,
biometric data, or other personal data from real human participants. Photorealistic,
pose-controlled human image generation carries a risk of misuse for impersonation or deceptive
synthetic media; such uses are explicitly not condoned. The dataset is intended for research in
human pose estimation, 3D human perception, and related computer vision applications.
## Licensing
Released for **non-commercial research use** upon publication. Full license terms, a datasheet
describing the dataset, intended use, and limitations will accompany the release.
## Citation
```bibtex
@inproceedings{poseatlas2027,
title = {PoseAtlas: A Multi-View Synthetic Human Dataset with Geometrically Grounded Poses for 3D Pose Estimation},
author = {PoseAtlas Authors},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
year = {2027}
}
```