Gtpbd Mm Eval

---
name: gtpbd-mm-eval
description: Evaluates multimodal terraced parcel extraction by measuring pixel-level segmentation accuracy, edge-level boundary recovery, and object-level structural consistency across image-only, image+text, and image+text+DEM input settings. Use when the user wants to benchmark on GTPBD-MM, or asks about evaluating this task. Reports mIoU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.12315
  bibtex_key: [REDACTED]
  confidence: high
---

# gtpbd-mm-eval

> GTPBD-MM: A Global Terraced Parcel and Boundary Dataset with Multi-Modality — Zhang et al. (2026) (arXiv:2604.12315, 2026)

## What this evaluates

Evaluates multimodal terraced parcel extraction by measuring pixel-level segmentation accuracy, edge-level boundary recovery, and object-level structural consistency across image-only, image+text, and image+text+DEM input settings.

## Datasets

- **GTPBD-MM** — total ?; splits: (unstated); repo https://github.com/Z-ZW-WXQ/GTPBD-MM

## Metrics

- `mIoU` **(primary)** — range: [0, 1]
  - Mean Intersection over Union. Computed as the average of IoU (intersection over union) across all classes or instances. IoU = TP / (TP + FP + FN).
- `F1` — range: [0, 1]
  - Harmonic mean of precision and recall. F1 = 2 * (Precision * Recall) / (Precision + Recall). Evaluates balance between false positives and false negatives.
- `ODS` — range: [0, 1]
  - Optimal Dataset Scale boundary metric. Measures boundary recovery quality by aggregating precision and recall across all images at the optimal threshold.
- `GTC` — range: [0, 1]
  - Geometric Total Consistency. Object-level metric reflecting structural consistency and geometric error at the instance level, penalizing merged or fragmented parcels.

## Input / output format

**Input**: High-resolution optical imagery, optional text descriptions, and optional Digital Elevation Model (DEM) data.

**Output**: Binary or multi-class segmentation mask delineating terraced parcels.

## Scoring recipe

```python
def compute_metrics(pred_mask, gt_mask):
    tp = (pred_mask & gt_mask).sum()
    fp = (pred_mask & ~gt_mask).sum()
    fn = (~pred_mask & gt_mask).sum()
    recall = tp / (tp + fn + 1e-6)
    precision = tp / (tp + fp + 1e-6)
    f1 = 2 * precision * recall / (precision + recall + 1e-6)
    iou = tp / (tp + fp + fn + 1e-6)
    # OIS/ODS require contour extraction & threshold sweep
    # GOC/GUC/GTC require instance matching & geometric error calculation
    return {'mIoU': iou, 'F1': f1, 'Recall': recall}
```

## Common pitfalls

- Metrics must be reported separately for each modality setting (Image-only, Image+Text, Image+Text+DEM) rather than aggregated.
- Object-level metrics (GOC, GUC, GTC) require instance-level matching and geometric error calculation, not just pixel-wise comparison.
- Edge metrics (OIS, ODS) depend on specific boundary extraction thresholds and contour matching tolerances detailed in Appendix C.

## Evidence (verbatim from paper)

> We adopt a three-level evaluation protocol. Pixel-level metrics include Recall, F1, OA, mIoU, and mAcc, which evaluate region-level segmentation quality. Edge-level metrics include OIS and ODS, which measure boundary recovery quality. Object-level metrics include GOC, GUC, and GTC, which reflect geometric error and structural consistency at the object level.

## Citation

```bibtex
@misc{zhang2026gtpbdmm,
  title={GTPBD-MM: A Global Terraced Parcel and Boundary Dataset with Multi-Modality},
  author={Zhang et al. (2026)},
  year={2026},
  note={arXiv:2604.12315}
}
```

- arXiv: 2604.12315