Loveda Seg Eval

---
name: loveda-seg-eval
description: Evaluates the effectiveness of a diffusion-based data augmentation method on mitigating long-tail bias and improving cross-domain generalization in remote-sensing semantic segmentation. It specifically probes whether synthetic label-image pairs can increase minority-class exposure while preserving domain realism and data distribution. Use when the user wants to benchmark on LoveDA, or asks about evaluating this task. Reports mIoU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2602.04749
  bibtex_key: [REDACTED]
  confidence: medium
---

# loveda-seg-eval

> Mitigating Long-Tail Bias via Prompt-Controlled Diffusion Augmentation — Wijenayake et al. (2026) (arXiv:2602.04749, 2026)

## What this evaluates

Evaluates the effectiveness of a diffusion-based data augmentation method on mitigating long-tail bias and improving cross-domain generalization in remote-sensing semantic segmentation. It specifically probes whether synthetic label-image pairs can increase minority-class exposure while preserving domain realism and data distribution.

## Datasets

- **LoveDA** — total ?; splits: test (-1)

## Metrics

- `mIoU` **(primary)** — range: [0, 1]
  - Mean Intersection over Union computed across all semantic classes. Calculated as the average of IoU (intersection over union of predicted and ground truth masks) for each class.

## Input / output format

**Input**: Remote-sensing satellite image paired with a pixel-level segmentation mask.

**Output**: Pixel-wise class label prediction for each image.

## Scoring recipe

```python
def compute_miou(predictions, ground_truth, num_classes):
    ious = []
    for c in range(num_classes):
        pred_c = (predictions == c)
        gt_c = (ground_truth == c)
        intersection = np.logical_and(pred_c, gt_c).sum()
        union = np.logical_or(pred_c, gt_c).sum()
        ious.append(intersection / union if union > 0 else 1.0)
    return np.mean(ious)
```

## Common pitfalls

- Reporting only head-class performance without stratifying by class frequency, which masks long-tail bias mitigation.
- Confusing in-domain evaluation (Urban→Urban or Rural→Rural) with cross-domain generalization (Urban→Rural or Rural→Urban).
- Assuming synthetic augmentation distorts the underlying data distribution, whereas the protocol explicitly verifies distribution preservation.

## Evidence (verbatim from paper)

> Table [I] shows that adding ratio-controlled synthetic pairs consistently improves segmentation across backbones, with the largest gains concentrated on minority and mid-tail classes rather than only the head classes. In-domain, mIoU increases for all models, with particularly strong improvements in agriculture, road, and water, indicating that synthesis mainly contributes context diversity for underrepresented semantics while respecting Urban/Rural style constraints.

## Citation

```bibtex
@misc{wijenayake2026mitigating,
  title={Mitigating Long-Tail Bias via Prompt-Controlled Diffusion Augmentation},
  author={Wijenayake et al. (2026)},
  year={2026},
  note={arXiv:2602.04749}
}
```

- arXiv: 2602.04749