Gridnethd Eval

---
name: gridnethd-eval
description: Evaluates 3D semantic segmentation capabilities for power line infrastructure using multi-modal LiDAR and image data. It probes a model's ability to accurately classify geometric and visual features into 11 distinct classes, including critical assets like pylons, cables, and insulators. Use when the user wants to benchmark on GridNet-HD, or asks about evaluating this task. Reports mIoU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2601.13052
  bibtex_key: [REDACTED]
  confidence: high
---

# gridnethd-eval

> GridNet-HD: A High-Resolution Multi-Modal Dataset for LiDAR-Image Fusion on Power Line Infrastructure — Carreaud et al. (2026) (arXiv:2601.13052, 2026)

## What this evaluates

Evaluates 3D semantic segmentation capabilities for power line infrastructure using multi-modal LiDAR and image data. It probes a model's ability to accurately classify geometric and visual features into 11 distinct classes, including critical assets like pylons, cables, and insulators.

## Datasets

- **GridNet-HD** — total 7694; splits: train (-1), val (-1), test (-1)

## Metrics

- `mIoU` **(primary)** — range: percent
  - Mean Intersection over Union computed across 11 semantic classes. IoU for a class is calculated as the number of correctly predicted pixels or points divided by the union of predicted and ground truth pixels or points for that class.

## Input / output format

**Input**: Co-georeferenced high-resolution LiDAR point clouds and corresponding images for power line infrastructure scenes.

**Output**: Per-point or per-voxel semantic class labels from a predefined 11-class ontology (e.g., Pylon, Conductor cable, Insulator, vegetation types, etc.).

## Scoring recipe

```python
def compute_miou(preds, gold, num_classes=11):
    ious = []
    for c in range(num_classes):
        tp = np.sum((preds == c) & (gold == c))
        fp = np.sum((preds == c) & (gold != c))
        fn = np.sum((preds != c) & (gold == c))
        iou = tp / (tp + fp + fn) if (tp + fp + fn) > 0 else 0.0
        ious.append(iou)
    return np.mean(ious) * 100
```

## Common pitfalls

- Models may report best single-run performance instead of averaging over 3 training runs with standard deviations.
- Test-Time Augmentation (TTA) and overlap strategies are applied to some baselines (PTv3, DITR) but not others, creating an unfair comparison if not explicitly noted.
- Class imbalance heavily impacts scores, with classes like 'Water' and 'Structural cable' showing drastically lower IoU than dominant classes like 'Pylon' or 'High vegetation'.

## Evidence (verbatim from paper)

> We report here the detailed per-class IoU scores on the test set for all baselines used in our study. Table 9 shows the average results over 3 training runs, including standard deviations for ImageVote, SPT, and Late Fusion. Table 10 reports the performance of the best model (highest mIoU) selected for each method.

## Citation

```bibtex
@misc{carreaud2026gridnethd,
  title={GridNet-HD: A High-Resolution Multi-Modal Dataset for LiDAR-Image Fusion on Power Line Infrastructure},
  author={Carreaud et al. (2026)},
  year={2026},
  note={arXiv:2601.13052}
}
```

- arXiv: 2601.13052