Stereo Video Deblurring Eval

---
name: stereo-video-deblurring-eval
description: Evaluates the ability of a stereo video deblurring algorithm to recover sharp frames from motion-blurred inputs, specifically testing its robustness to spatially-variant blur caused by independent 3D object motion and non-planar surfaces. Use when the user wants to benchmark on Custom synthetic raytraced & real stereo captures, or asks about evaluating this task. Reports PSNR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1607.08421
  bibtex_key: sellent2016stereo
  confidence: high
---

# stereo-video-deblurring-eval

> Stereo Video Deblurring — Sellent et al. (2016) (arXiv:1607.08421, 2016)

## What this evaluates

Evaluates the ability of a stereo video deblurring algorithm to recover sharp frames from motion-blurred inputs, specifically testing its robustness to spatially-variant blur caused by independent 3D object motion and non-planar surfaces.

## Datasets

- **Custom synthetic raytraced & real stereo captures** — total ?; splits: test (-1)

## Metrics

- `PSNR` **(primary)** — range: other
  - Peak Signal-to-Noise Ratio computed between the deblurred output and the ground-truth sharp reference frame. Standard logarithmic ratio of max pixel intensity squared to mean squared error.
- `AEP` — range: other
  - Average Endpoint Error measuring the mean Euclidean distance between estimated and ground-truth 2D motion vectors.
- `ADE` — range: other
  - Average Disparity Error measuring the mean absolute difference between estimated and ground-truth disparity maps.

## Input / output format

**Input**: Blurred left and right stereo video frames (640x480), camera calibration parameters, and exposure times. For synthetic data, ground-truth sharp frames and known 3D motion/disparity are provided.

**Output**: Deblurred left and right stereo frames, estimated scene flow/disparity maps, and motion boundary masks.

## Scoring recipe

```python
def compute_psnr(gt, pred):
    mse = np.mean((gt - pred) ** 2)
    if mse == 0:
        return 100.0
    max_pixel = 255.0
    return 10 * np.log10((max_pixel ** 2) / mse)

def compute_aep(gt_flow, pred_flow):
    return np.mean(np.sqrt(np.sum((gt_flow - pred_flow) ** 2, axis=-1)))

def compute_ade(gt_disp, pred_disp):
    return np.mean(np.abs(gt_disp - pred_disp))
```

## Common pitfalls

- Using 2D optical flow instead of 3D scene flow/homographies for non-fronto-parallel motion leads to significant accuracy drops and ringing artifacts.
- Disparity estimation failures on complex, non-planar objects (e.g., 'apples') can degrade homography-based deblurring unless motion discontinuities are explicitly masked and downweighted.
- Assuming constant linear velocity for 3D accelerated motion (like forward translation) causes blur kernel misalignment, which homography-based models handle better.

## Evidence (verbatim from paper)

> Table 2 shows the peak-signal-to-noise-ratio (PSNR) of the deblurred images from the different methods. We observe that the PSNR of our homography-based stereo video deblurring outperforms the results of deblurring with ground truth 2D displacement in all cases of non-fronto-parallel motion.

## Citation

```bibtex
@misc{sellent2016stereo,
  title={Stereo Video Deblurring},
  author={Sellent et al. (2016)},
  year={2016},
  note={arXiv:1607.08421}
}
```

- arXiv: 1607.08421