pass-worker-detection-yolov10s

YOLOv10-s worker detection checkpoints trained on real construction-site imagery (MOCS + SODA) and augmented with increasing amounts of synthetic data from two generation pipelines — 3D reconstruction and LiDAR — across two construction sites (small_site and large_site).

The collection studies how synthetic-data source and volume affect detection accuracy, sweeping synthetic augmentation from 2.5k to 17.5k images in 2.5k increments.

Intended use

These are site-adaptive detectors tuned to specific construction-site domains and the MOCS+SODA label space. They are intended for research and for deployment on scenes matching the training domains. They are not intended to generalize to arbitrary object-detection tasks, unrelated domains, or open-world imagery. Validate on your own data before any operational use.

Repository structure

pass-worker-detection-yolov10s/
├── baseline/
│   └── real_only_mocs_soda.pt          # real-only MOCS+SODA baseline
├── small_site/                         # small site
│   ├── recon/                          # + 3D-reconstruction-based synthetic data
│   │   ├── real_2.5k.pt
│   │   ├── real_5k.pt
│   │   ├── real_7.5k.pt
│   │   ├── real_10k.pt
│   │   ├── real_12.5k.pt
│   │   ├── real_15k.pt
│   │   └── real_17.5k.pt
│   └── lidar/                          # + LiDAR-based synthetic data
│       ├── real_2.5k.pt
│       ├── real_5k.pt
│       ├── real_7.5k.pt
│       ├── real_10k.pt
│       ├── real_12.5k.pt
│       ├── real_15k.pt
│       └── real_17.5k.pt
└── large_site/                         # large site
    ├── recon/                          # + 3D-reconstruction-based synthetic data
    │   ├── real_2.5k.pt
    │   ├── real_5k.pt
    │   ├── real_7.5k.pt
    │   ├── real_10k.pt
    │   ├── real_12.5k.pt
    │   ├── real_15k.pt
    │   └── real_17.5k.pt
    └── lidar/                          # + LiDAR-based synthetic data
        ├── real_2.5k.pt
        ├── real_5k.pt
        ├── real_7.5k.pt
        ├── real_10k.pt
        ├── real_12.5k.pt
        ├── real_15k.pt
        └── real_17.5k.pt

Training configuration

All models were trained with the same recipe:

• Model: YOLOv10-s (yolov10s.yaml)
• Epochs: 800
• Batch size: 64
• Image size: 640
• lr0: 0.001
• lrf: 0.01

Models

Every checkpoint is a YOLOv10-s model (best.pt). AP@50-95 values are the results of evaluating each model on real test sets collected and manually labeled from the small and large sites.