ayushgupta7777 / safetyvision-yolov8

SafetyVision YOLOv8n — PPE Detection

YOLOv8n fine-tuned for Personal Protective Equipment (PPE) detection at industrial worksites. Backbone model for SafetyVision , an open-source AI workplace safety monitor.

Model description

YOLOv8n fine-tuned for 13-class PPE detection covering hard hats, safety vests, goggles, gloves, masks, fall harnesses, their "missing/no" violation counterparts, fall detection, and a Person class.

• Base model: Ultralytics YOLOv8n (AGPL-3.0)
• Output: 17 channels × 8400 anchor predictions → NMS → bounding boxes + class labels + confidence scores
• Use it for: flagging likely PPE violations in static images and short video clips for human review
• Do not use it for: automated disciplinary action, medical/clinical PPE, food safety, hazmat suits, or any standalone enforcement decision

Classes

Training data

PPE-Combined v1 (Roboflow Universe, forked from s-workspace-cjeuu/ppe-combined-9bprl ).

Training procedure

• Hardware: Kaggle Notebooks, 2× NVIDIA Tesla T4 (15.6 GB VRAM each)
• Framework: Ultralytics 8.3.40, PyTorch 2.10.0 + CUDA 12.8
• Epochs: 100
• Batch size: 32
• Image size: 640×640
• Optimizer: SGD (lr0=0.01, momentum=0.937, weight_decay=0.0005, default ultralytics schedule)
• Wall time: ~15 hours across two Kaggle Save Versions (12-hour session cap forced a mid-run resume at epoch 82)
• Resume: Save Version 1 crashed at epoch 82.5/100 on the Kaggle 12-hour cap. Save Version 2 resumed from the epoch82.pt checkpoint via wandb.init(id="9nctv2ai", resume="must", settings=wandb.Settings(init_timeout=300)) and completed epochs 83–100 cleanly.

Experiment tracking

• W&B run (public): https://wandb.ai/agcr7jw-vellore-institute-of-technology/safetyvision/runs/9nctv2ai
  • Epochs 1–82: full charts available
  • Epochs 83–100: W&B callback didn't fire after resume (known ultralytics+wandb sharp edge). Canonical metrics for all 100 epochs are in model/yolov8n-ppe-v1/results.csv in the repo.
• MLflow: local file-based tracking committed at mlruns/ , run ID f1932e539038417dad6db757affd50e6 .

Evaluation

Headline metrics (held-out test split, 5,148 images)

Test split is slightly better than validation — clean generalization signal, no overfitting.

Per-class test metrics

Curves and confusion matrices for both splits are committed in model/yolov8n-ppe-v1/ .

Inference performance

• GPU (Tesla T4): preprocess 0.2 ms + inference 2.9 ms + postprocess 1.0 ms = ~4 ms/image total

Intended use

Pre-screening tool to assist human workplace safety officers by surfacing likely PPE violations in images and short video clips for human review. Designed for:

• Construction sites
• Warehouses
• Manufacturing floors
• Pre-shift safety walkthroughs

Not a replacement for human judgment. Predictions must be reviewed by qualified safety personnel before any disciplinary, compliance, or insurance action.

Out of scope

• Medical/clinical settings (different PPE: gowns, N95 fit testing, sterile gloves)
• Food processing (hairnets, beard guards, lab coats not represented)
• Chemical/hazmat operations (full-face respirators, encapsulating suits not represented)
• Drone or overhead camera angles (training data is ground-level/eye-level)
• Crowded scenes with heavy mutual occlusion
• Real-time alerting where missing a violation is unacceptable

Failure modes

Documented from training data review and observed test errors:

• Low light or high glare — Confidence drops sharply; expect both false positives and false negatives.
• Partial occlusion — Workers partially behind machinery or other workers may have PPE missed.
• Unusual PPE colors — Training data skews to high-vis yellow/orange/lime vests and standard white/yellow hard hats. Rare colors (blue, black) may go undetected.
• Small workers (<50 px height) — Distant figures often missed entirely.
• Fast motion in video — Motion blur causes missed frames. Mitigate by aggregating across multiple frames per scene rather than relying on any single frame.
• No_Harness class — Severely underrepresented in training and test data (1 test instance, 276 val instances). Effectively unusable as a detector until training data is augmented. Do not rely on this class for fall-arrest compliance.
• NO-Safety Vest class — Weakest violation class (test mAP 0.395). High false-negative rate.

Bias and limitations

• Training data over-represents Western construction/industrial sites; demographics and PPE conventions in South/Southeast Asia, Africa, and the Middle East may be underrepresented.
• Heavily skewed toward male-presenting workers in training imagery.
• The Person class inherits biases from the underlying YOLOv8 COCO pretraining.
• Indoor warehouse lighting is overrepresented; bright outdoor sun and underground/tunnel environments may degrade performance.
• 13 classes is a fixed taxonomy — site-specific PPE (e.g., arc-flash hoods, cut-resistant sleeves) will not be detected.

Files

Usage

PyTorch (ultralytics)

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

weights = hf_hub_download(repo_id="ayushgupta7777/safetyvision-yolov8", filename="best.pt")
model = YOLO(weights)
results = model("worksite_image.jpg")
results[0].show()

ONNX Runtime (CPU-friendly, used in AWS Lambda)

import cv2, numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download

# Both .onnx and .onnx.data must be in the same directory
hf_hub_download(repo_id="ayushgupta7777/safetyvision-yolov8", filename="best.onnx.data")
onnx_path = hf_hub_download(repo_id="ayushgupta7777/safetyvision-yolov8", filename="best.onnx")

session = ort.InferenceSession(onnx_path)
img = cv2.imread("worksite_image.jpg")
img = cv2.resize(img, (640, 640))
inp = img.transpose(2, 0, 1)[None].astype(np.float32) / 255.0
outputs = session.run(None, {"images": inp})
# outputs[0] shape: (1, 17, 8400) — apply your own NMS for final boxes

License

• Model weights: AGPL-3.0 (inherited from Ultralytics YOLOv8 base model)
• SafetyVision repository code: see LICENSE in the project repo.

Citation

@software{safetyvision_2026,
  author = {Gupta, Ayush},
  title  = {SafetyVision: Open-Source AI Workplace Safety Monitor},
  year   = {2026},
  url    = {https://github.com/ayushgupta07xx/SafetyVision}
}

Acknowledgements

• Ultralytics for YOLOv8 and the training framework
• Roboflow Universe and the PPE-Combined dataset maintainers
• OSHA for the public-domain regulation corpus used in incident report generation
• Kaggle Notebooks for free 2× T4 GPU training