Biwi Head Pose Eval

---
name: biwi-head-pose-eval
description: Evaluates monocular head pose estimation accuracy by predicting 6DoF rotation (yaw, pitch, roll) from RGB images, comparing absolute single-image regression against relative two-view transformation prediction. Use when the user wants to benchmark on BIWI Kinect Head Pose Database, or asks about evaluating this task. Reports MAE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.10106
  bibtex_key: vasileiou2026vggthpe
  confidence: high
---

# biwi-head-pose-eval

> VGGT-HPE: Reframing Head Pose Estimation as Relative Pose Prediction — Vasileiou et al. (2026) (arXiv:2604.10106, 2026)

## What this evaluates

Evaluates monocular head pose estimation accuracy by predicting 6DoF rotation (yaw, pitch, roll) from RGB images, comparing absolute single-image regression against relative two-view transformation prediction.

## Datasets

- **BIWI Kinect Head Pose Database** — total ?; splits: test (-1)

## Metrics

- `MAE` **(primary)** — range: degrees
  - Mean Absolute Error averaged across yaw, pitch, and roll rotation angles. Computed as the average of the absolute differences between predicted and ground-truth angles in degrees.

## Input / output format

**Input**: Single RGB image (absolute mode) or a pair of RGB images consisting of a target frame and an anchor frame with known ground-truth pose (relative mode). Faces are cropped using MTCNN.

**Output**: Predicted 6DoF head pose rotation specified as yaw, pitch, and roll angles in degrees.

## Scoring recipe

```python
def compute_mae(pred_angles, gt_angles):
    # pred_angles, gt_angles: arrays of shape (N, 3) for [yaw, pitch, roll]
    abs_errors = np.abs(pred_angles - gt_angles)
    mae_per_axis = abs_errors.mean(axis=0)
    overall_mae = mae_per_axis.mean()
    return overall_mae
```

## Common pitfalls

- Face detection and cropping strategy must be standardized (MTCNN used here) to ensure fair comparison across baselines.
- Ground-truth poses are originally in the depth-camera frame and must be transformed to the RGB-camera frame using per-subject calibration parameters.
- Relative model performance depends on anchor frame selection; using the first frame of a sequence as a fixed anchor requires only one ground-truth pose per subject.

## Evidence (verbatim from paper)

> We report mean absolute error (MAE) in degrees for yaw, pitch, and roll, along with the overall MAE averaged across the three axes.

## Citation

```bibtex
@misc{vasileiou2026vggthpe,
  title={VGGT-HPE: Reframing Head Pose Estimation as Relative Pose Prediction},
  author={Vasileiou et al. (2026)},
  year={2026},
  note={arXiv:2604.10106}
}
```

- arXiv: 2604.10106