SinaDavari/pass-worker-detection-test-sets

PASS — Real Annotated Worker-Detection Test Sets

Real-world, manually annotated construction-worker (person) detection test sets from two construction sites — a small site and a large site.

These are the held-out evaluation sets used to benchmark the site-adaptive YOLOv10-s detectors in the PASS (Passive-sensing Site Synthesis) study. Every model in the companion model repo SinaDavari/pass-worker-detection-yolov10s was evaluated on exactly these images. They are intended for evaluation only (no train/val split is provided).

Dataset summary

Site	Split	Images	Boxes	Negatives	Classes	Image size
small_site	test	1,061	1,672	0	1 (person)	1280×720
large_site	test	752	1,131	0	1 (person)	1280×720

• Task: single-class 2D object detection (worker / person).
• Annotations: manually labeled, axis-aligned bounding boxes.
• Provenance: frames captured on real construction sites, then hand-annotated.

Structure

Native Ultralytics YOLO layout — one directory per site, plus a ready-to-use data config:

pass-worker-detection-test-sets/
├── small_site/
│   ├── images/test/*.jpg      # 1,061 images
│   └── labels/test/*.txt      # YOLO: <class> <cx> <cy> <w> <h> (normalized)
├── large_site/
│   ├── images/test/*.jpg      # 752 images
│   └── labels/test/*.txt
├── small_site.yaml            # ultralytics data config
└── large_site.yaml

Labels are YOLO format, class 0 = person, coordinates normalized to [0, 1].

Usage

Download the repo, then evaluate any checkpoint with Ultralytics / YOLOv10:

# 1. fetch the datasets
huggingface-cli download SinaDavari/pass-worker-detection-test-sets \
    --repo-type dataset --local-dir pass-test-sets

# 2. fetch a model checkpoint from the companion model repo
huggingface-cli download SinaDavari/pass-worker-detection-yolov10s \
    small_site/recon/real_15k.pt --local-dir models

# 3. evaluate (run from inside pass-test-sets/)
yolo val model=../models/small_site/recon/real_15k.pt data=small_site.yaml

Reported metric in the paper is mAP@50–95 (COCO-style, IoU 0.50→0.95).

Intended use and limitations

• Intended use: benchmarking construction worker/person detectors, especially for the site-adaptation study in PASS.
• Site-specific: the two sites differ in scene and scale; models tuned to one site are not expected to transfer to the other or to arbitrary domains.
• Evaluation only: no training/validation partition is provided.

License

Released under CC BY-NC 4.0 (attribution, non-commercial). This imagery was collected on real construction sites and contains people; please use it responsibly and for research purposes. (Maintainer: confirm/adjust this license if a different one applies.)

Citation

Companion to the paper "PASS: Passive-sensing Site Synthesis" (submitted to Automation in Construction). Code: github.com/SinaDavari/PaSS. Full citation to be added upon acceptance.