Viewdelta Scd Eval

---
name: viewdelta-scd-eval
description: This evaluation probes a model's ability to perform scene change detection conditioned on natural language prompts. It measures how well the model distinguishes relevant semantic changes from nuisance variations across diverse domains (street-view, satellite, indoor) and handles viewpoint misalignments. Use when the user wants to benchmark on CSeg, PSCD, SYSU-CD, VL-CMU-CD, or asks about evaluating this task. Reports IoU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2412.07612
  bibtex_key: [REDACTED]
  confidence: high
---

# viewdelta-scd-eval

> ViewDelta: Scaling Scene Change Detection through Text-Conditioning — Varghese et al. (2024) (arXiv:2412.07612, 2024)

## What this evaluates

This evaluation probes a model's ability to perform scene change detection conditioned on natural language prompts. It measures how well the model distinguishes relevant semantic changes from nuisance variations across diverse domains (street-view, satellite, indoor) and handles viewpoint misalignments.

## Datasets

- **CSeg** — total 500000; splits: test (-1)
- **PSCD** — total ?; splits: test (-1)
- **SYSU-CD** — total ?; splits: test (-1)
- **VL-CMU-CD** — total ?; splits: test (-1)

## Metrics

- `IoU` **(primary)** — range: percent
  - Intersection over Union: IoU = TP / (TP + FP + FN). Measures the overlap between predicted and ground truth change masks.
- `F1 score` — range: percent
  - Harmonic mean of Precision and Recall: F1 = 2 * (Precision * Recall) / (Precision + Recall).
- `Recall` — range: percent
  - True Positive Rate: Recall = TP / (TP + FN). Measures the proportion of actual changes correctly detected.
- `Precision` — range: percent
  - Positive Predictive Value: Precision = TP / (TP + FP). Measures the proportion of predicted changes that are correct.

## Input / output format

**Input**: Bi-temporal image pairs accompanied by a natural language text prompt specifying the types of changes to detect.

**Output**: A segmentation mask (binary or multi-class) indicating changed regions, generated per image pair.

## Scoring recipe

```python
def compute_metrics(pred_mask, gt_mask):
    pred = pred_mask.flatten()
    gt = gt_mask.flatten()
    tp = np.sum((pred == 1) & (gt == 1))
    fp = np.sum((pred == 1) & (gt == 0))
    fn = np.sum((pred == 0) & (gt == 1))
    iou = tp / (tp + fp + fn + 1e-8)
    precision = tp / (tp + fp + 1e-8)
    recall = tp / (tp + fn + 1e-8)
    f1 = 2 * precision * recall / (precision + recall + 1e-8)
    return {'IoU': iou, 'F1': f1, 'Recall': recall, 'Precision': precision}
```

## Common pitfalls

- Prompting strategies are dataset-specific and must be adapted: class names for PSCD, comprehensive descriptions for SYSU-CD, or 'all' for unaligned variants.
- VL-CMU-CD contains label noise; using prompts that match consistent human labels is required for fair evaluation.
- Unaligned variants (Diff-1/Diff-2) test robustness to viewpoint shifts rather than standard change detection, requiring models to avoid spatial alignment assumptions.

## Evidence (verbatim from paper)

> We report standard change detection metrics including Intersection over Union (IoU), F1 score, Recall, and Precision. For each dataset, we adopt task-appropriate prompting strategies: dataset-specific class names for PSCD, a comprehensive change description for SYSU-CD, and a description of object categories for VL-CMU-CD.

## Citation

```bibtex
@misc{varghese2024viewdelta,
  title={ViewDelta: Scaling Scene Change Detection through Text-Conditioning},
  author={Varghese et al. (2024)},
  year={2024},
  note={arXiv:2412.07612}
}
```

- arXiv: 2412.07612