Person Detection Eval

---
name: person-detection-eval
description: Evaluates person and body part detection accuracy using standard object detection metrics, while assessing a self-monitoring framework's ability to reduce false negatives and false positives through part-based plausibility checks. Use when the user wants to benchmark on DensePose, MS-COCO, Pascal VOC, or asks about evaluating this task. Reports AP@0.5.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2307.04533
  bibtex_key: schwaiger2023preventing
  confidence: high
---

# person-detection-eval

> Preventing Errors in Person Detection: A Part-Based Self-Monitoring Framework — Schwaiger et al. (2023) (arXiv:2307.04533, 2023)

## What this evaluates

Evaluates person and body part detection accuracy using standard object detection metrics, while assessing a self-monitoring framework's ability to reduce false negatives and false positives through part-based plausibility checks.

## Datasets

- **DensePose** — total 15698; splits: train (13483), val (2215)
- **MS-COCO** — total 11691; splits: val (11691)
- **Pascal VOC** — total 2971; splits: trainval (2971)

## Metrics

- `mAP` — range: [0, 1]
  - Mean Average Precision across all classes, computed by averaging the Area Under the Precision-Recall curve for each class.
- `AP@0.5` **(primary)** — range: [0, 1]
  - Average Precision computed using a fixed Intersection over Union (IoU) threshold of 0.5 between predicted and ground-truth bounding boxes.

## Input / output format

**Input**: RGB images containing persons and body parts.

**Output**: Bounding boxes and class labels for 'person' and 8 body parts (Torso, Hand, Foot, Upper Leg, Lower Leg, Upper Arm, Lower Arm, Head).

## Scoring recipe

```python
1. Match predicted boxes to ground-truth boxes using IoU >= 0.5.
2. Classify matches as TP, unmatched predictions as FP, unmatched GT as FN.
3. Compute Precision-Recall curve per class.
4. Calculate AP per class by integrating the PR curve.
5. Average AP across classes to get mAP. Report AP@0.5 as the headline metric.
```

## Common pitfalls

- IoU threshold is fixed at 0.5, not the standard COCO 0.5:0.95 range.
- Evaluation images are filtered to exclude persons with bounding box area < 2247 pixels^2.
- Cross-dataset evaluation uses the Pascal VOC trainval split, not the standard test split.

## Evidence (verbatim from paper)

> The performance for class person in terms of mAP and Average Precision (AP) with an IOU threshold of 0.5 is presented for each model in Table [I]... To determine the detection sets derived from the ground-truth annotations D_TP_gt, D_FP_gt, and D_FN_gt, we set the value of the IOU threshold to 0.5.

## Citation

```bibtex
@misc{schwaiger2023preventing,
  title={Preventing Errors in Person Detection: A Part-Based Self-Monitoring Framework},
  author={Schwaiger et al. (2023)},
  year={2023},
  note={arXiv:2307.04533}
}
```

- arXiv: 2307.04533