Hand Avatar Personalization Eval

---
name: hand-avatar-personalization-eval
description: Evaluates the ability of a model to personalize a 3D hand avatar from a single RGB image and render it under novel poses and lighting conditions. It probes physically-based rendering accuracy, material/albedo recovery, and relighting generalization. Use when the user wants to benchmark on InterHand2.6M, HARP relit, or asks about evaluating this task. Reports PSNR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2511.05403
  bibtex_key: fan2025palm
  confidence: high
---

# hand-avatar-personalization-eval

> PALM: A Dataset and Baseline for Learning Multi-subject Hand Prior — Fan et al. (2025) (arXiv:2511.05403, 2025)

## What this evaluates

Evaluates the ability of a model to personalize a 3D hand avatar from a single RGB image and render it under novel poses and lighting conditions. It probes physically-based rendering accuracy, material/albedo recovery, and relighting generalization.

## Datasets

- **InterHand2.6M** — total ?; splits: test (240)
- **HARP relit** — total ?; splits: train (-1), eval (-1)

## Metrics

- `PSNR` **(primary)** — range: dB
  - Peak Signal-to-Noise Ratio, calculated as 10 * log10(MAX^2 / MSE) where MAX is the maximum pixel value (typically 255 for 8-bit images) and MSE is the mean squared error between predicted and ground truth images.
- `SSIM` — range: [0, 1]
  - Structural Similarity Index, measuring luminance, contrast, and structure between two images to assess perceptual similarity.
- `LPIPS` — range: [0, 1]
  - Learned Perceptual Image Patch Similarity, computed using a pre-trained deep network to measure perceptual distance between images.

## Input / output format

**Input**: Single RGB image of a hand (used to train a per-sequence personalized model) and ground-truth reference images for evaluation.

**Output**: Rendered RGB image of the personalized hand avatar under a novel pose and/or environment lighting.

## Scoring recipe

```python
def evaluate(pred_rgb, gt_rgb):
    mse = np.mean((pred_rgb - gt_rgb) ** 2)
    psnr = 10 * np.log10(255**2 / mse)
    ssim = compute_ssim(pred_rgb, gt_rgb)
    lpips = compute_lpips(pred_rgb, gt_rgb)
    return {'PSNR': psnr, 'SSIM': ssim, 'LPIPS': lpips}
```

## Common pitfalls

- Per-sequence personalization: The protocol requires training a separate personalized model for each sequence using only its first frame, rather than optimizing a single global model.
- Novel environment relighting: Evaluation explicitly uses different environment maps for training and testing, meaning models must generalize lighting rather than memorize it.
- Metric interpretation: LPIPS and SSIM are perceptual/structural metrics that may not perfectly align with visual fidelity in complex PBR materials, so qualitative inspection is recommended alongside quantitative scores.

## Evidence (verbatim from paper)

> We evaluate our baseline on the task of hand avatar personalization and relighting from a single RGB image using three different datasets. Metric and baseline details are in SupMat. Table 2: InterHand2.6M dataset evaluation. Comparison of methods on single-image personalization task using PSNR, SSIM, and LPIPS metrics.

## Citation

```bibtex
@misc{fan2025palm,
  title={PALM: A Dataset and Baseline for Learning Multi-subject Hand Prior},
  author={Fan et al. (2025)},
  year={2025},
  note={arXiv:2511.05403}
}
```

- arXiv: 2511.05403