Persense D Eval

---
name: persense-d-eval
description: Evaluates training-free, one-shot instance segmentation in dense, cluttered, and occluded scenes. It probes the model's ability to localize and segment specific target instances using few exemplars and point prompts, while handling high object density and overlapping objects. Use when the user wants to benchmark on PerSense-D, COCO-20i, COCO-20d, LVIS-92i, LVIS-92d, or asks about evaluating this task. Reports mIoU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2508.14660
  bibtex_key: siddiqui2025persense
  confidence: high
---

# persense-d-eval

> Towards PerSense++: Advancing Training-Free Personalized Instance Segmentation in Dense Images — Siddiqui et al. (2025) (arXiv:2508.14660, 2025)

## What this evaluates

Evaluates training-free, one-shot instance segmentation in dense, cluttered, and occluded scenes. It probes the model's ability to localize and segment specific target instances using few exemplars and point prompts, while handling high object density and overlapping objects.

## Datasets

- **PerSense-D** — total 717; splits: test (717)
- **COCO-20i** — total 41000; splits: test (41000)
- **COCO-20d** — total 9184; splits: test (9184)
- **LVIS-92i** — total 20000; splits: test (20000)
- **LVIS-92d** — total 7204; splits: test (7204)

## Metrics

- `mIoU` **(primary)** — range: [0, 1]
  - Mean Intersection over Union: the average of the IoU (intersection over union) computed across all object classes or instances in the dataset.
- `MAE` — range: other
  - Mean Absolute Error: the average absolute difference between the predicted density map values and the ground truth density map values across all pixels.
- `RMSE` — range: other
  - Root Mean Square Error: the square root of the average of squared differences between predicted and ground truth density map values.

## Input / output format

**Input**: A query image, one or more exemplar/reference images of the target instance, and point prompts (coordinates) indicating instance locations.

**Output**: Binary instance segmentation masks for each queried instance.

## Scoring recipe

```python
def compute_miou(pred_masks, gt_masks):
    ious = []
    for pred, gt in zip(pred_masks, gt_masks):
        intersection = np.logical_and(pred, gt).sum()
        union = np.logical_or(pred, gt).sum()
        if union == 0:
            ious.append(1.0)
        else:
            ious.append(intersection / union)
    return np.mean(ious)
```

## Common pitfalls

- COCO/LVIS dense ground truths are coarse blob-like masks rather than instance-level annotations, which underrepresents fine-grained instance segmentation performance.
- Density map generation is inefficient on sparse images with few instances, making traditional detection methods more suitable for low-density scenes.

## Evidence (verbatim from paper)

> We use standard evaluation metric of mean Intersection over Union (mIoU) for evaluating segmentation performance. To assess the accuracy of the generated DMs, particularly when comparing PerSense with PerSense++, we report Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), following established object counting protocols. No training is involved in any of our experiments.

## Citation

```bibtex
@misc{siddiqui2025persense,
  title={Towards PerSense++: Advancing Training-Free Personalized Instance Segmentation in Dense Images},
  author={Siddiqui et al. (2025)},
  year={2025},
  note={arXiv:2508.14660}
}
```

- arXiv: 2508.14660