SafetyVision YOLOv8 — PPE Detection (v1 nano · v2 small) YOLOv8 fine tuned for Personal Protective Equipment (PPE) detection at industrial worksites. Backbone model for SafetyVision, an open source AI workplace safety monitor. This repo hosts two versions : v2 (current, production) — YOLOv8s, trained on 80k images with Albumentations augmentation. Weights at v2/ . v1 (original) — YOLOv8n, trained on 58k images. Weights at the repo root, kept for reproducibility and the v1→v2 comparison. Headline metric (v2, held out test) Value Test mAP@0.5 (imgsz 896) 0.766 Test mAP@0.5 (imgsz 640) 0.754 Deployed ONNX mAP@0.5 (imgsz 640) 0.738 Test mAP@0.5:0.95 (imgsz 896) 0.487 Validation mAP@0.5 0.787 Parameters 11,130,615 (~11.1M) FLOPs 28.5 GFLOPs Honest note on the target. The Phase 2 goal was mAP@0.5 ≥ 0.78 on the held out test split. Validation cleared it (0.787); the held out test came in at 0.766 (imgsz 896) — short of 0.78 by 0.014. We report the test number as the headline generalization figure rather than leading with the higher validation value. See Evaluation. What's new in v2 (v1 → v2) Aspect v1 (YOLOv8n) v2 (YOLOv8s) Backbone nano small Parameters ~3.0M ~11.1M Training images 57,90…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy