// the find
njvisionpower/Safety-Helmet-Wearing-Dataset
Safety helmet wearing detect dataset, with pretrained model
A labeled dataset of 7581 images for detecting safety helmet compliance in industrial/construction settings, with pretrained YOLO models via MXNet GluonCV. Useful for anyone building CV pipelines for workplace safety monitoring. The dataset itself is the main product — the training code is thin scaffolding around GluonCV.
Pascal VOC format means it drops into most object detection pipelines without conversion; three model sizes (darknet53, mobilenet 1.0, mobilenet 0.25) give you a tradeoff between accuracy (88.5 mAP) and deployment weight; negative examples sourced from SCUT-HEAD give the 'no helmet' class real diversity rather than synthetic negatives; dataset is reasonably large for a domain-specific safety task with 9K positive and 111K negative annotations.
Frozen since December 2019 — GluonCV is effectively abandoned, MXNet is dead as a mainstream framework, so you'd need to port to PyTorch/ONNX yourself before using this in anything modern; dataset is hosted exclusively on BaiduDrive/GoogleDrive with no versioning, so if those links die, the repo is just demo images; only two classes ('hat' and 'person') means no distinction between hard hat types, partial compliance, or other PPE — too coarse for real safety audits; no license information on the dataset itself (only the code has MIT), which is a problem for commercial use.