Crosspoint Bench Eval

---
name: crosspoint-bench-eval
description: Evaluates Vision-Language Models' ability to perform precise point-level geometric correspondence across multiple viewpoints. It probes fine-grained spatial grounding, visibility reasoning, cross-view correspondence judgment, and continuous 2D coordinate pointing. Use when the user wants to benchmark on CrossPoint-Bench, or asks about evaluating this task. Reports average accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2512.04686
  bibtex_key: wang2025crosspoint
  confidence: high
---

# crosspoint-bench-eval

> Towards Cross-View Point Correspondence in Vision-Language Models — Wang et al. (2025) (arXiv:2512.04686, 2025)

## What this evaluates

Evaluates Vision-Language Models' ability to perform precise point-level geometric correspondence across multiple viewpoints. It probes fine-grained spatial grounding, visibility reasoning, cross-view correspondence judgment, and continuous 2D coordinate pointing.

## Datasets

- **CrossPoint-Bench** — total ?; splits: test (-1); repo https://github.com/WangYipu2002/CrossPoint

## Metrics

- `average accuracy` **(primary)** — range: [0, 1]
  - Proportion of correctly answered multiple-choice questions across all evaluated instances.
- `in-mask hit rate` — range: [0, 1]
  - Fraction of predicted 2D coordinates that fall within the ground-truth mask for pointing tasks.

## Input / output format

**Input**: Multi-view images paired with textual prompts for multiple-choice questions or point-targeting instructions.

**Output**: Selected option index/text for multiple-choice tasks; predicted (x, y) 2D coordinates for pointing tasks.

## Scoring recipe

```python
def compute_metrics(predictions, golds, task_types):
    acc_scores = []
    hit_scores = []
    for pred, gold, task in zip(predictions, golds, task_types):
        if task == 'multiple-choice':
            acc_scores.append(1.0 if pred == gold else 0.0)
        elif task == 'pointing':
            hit_scores.append(1.0 if gold_mask.contains(pred) else 0.0)
    return {
        'average_accuracy': sum(acc_scores) / len(acc_scores) if acc_scores else 0,
        'in_mask_hit_rate': sum(hit_scores) / len(hit_scores) if hit_scores else 0
    }
```

## Common pitfalls

- Frame transfer failure: models often reason within the source view's coordinate system instead of mapping to the target view.
- Spatial reconstruction failure: models struggle to integrate occlusion and relative layout into a consistent 3D representation.
- Semantic-point decoupling: models may correctly identify the target object semantically but fail to align it to the precise pixel location.

## Evidence (verbatim from paper)

> On CrossPoint-Bench, we adopt two complementary metrics: multiple-choice tasks use average accuracy, while pointing tasks use in-mask hit rate, which measures whether the predicted 2D coordinate falls within the ground-truth mask.

## Citation

```bibtex
@misc{wang2025crosspoint,
  title={Towards Cross-View Point Correspondence in Vision-Language Models},
  author={Wang et al. (2025)},
  year={2025},
  note={arXiv:2512.04686}
}
```

- arXiv: 2512.04686