fiacecson20/cctv-ai-ppe

cctv-ai · ppe · YOLOv8n round-1

YOLOv8n trained on CPU for the CCTV-AI project. Round-1 baseline — see docs/TRAINING_REPORT.md in the upstream repo for full eval details.

Files

File	Purpose
best.onnx	ONNX export (opset 12), drop-in for onnxruntime / OpenVINO / TensorRT.
args.yaml	Full training arguments (reproducibility).
results.csv	Per-epoch metrics.
confusion_matrix.png	Validation confusion matrix.
BoxPR_curve.png	Precision-recall curve (per class).
BoxF1_curve.png	F1 curve (per class).
results.png	Training curves (loss + metrics across epochs).
val_batch0_pred.jpg / val_batch0_labels.jpg	Sample predictions vs. ground truth.

The .pt weights live inside the upstream repo at weights/ppe-yolov8n-v0.pt via Git LFS — .onnx lives here because it exceeds the 10 MB LFS threshold we picked for in-repo storage.

Eval (val split)

• Classes: helmet, head, vest
• Overall val mAP@0.5: 0.876
• Inference: ~44 ms / frame on 2-core AMD EPYC CPU (≈22 FPS)

class	precision	recall	mAP@0.5
helmet	0.939	0.806	0.905
head	0.860	0.785	0.847
vest	0.000	0.000	0.000

Training config

• Model: yolov8n.pt (Ultralytics)
• imgsz=320, batch=8, epochs=12, patience=4, cos_lr=true
• Optimizer: AdamW (auto)
• Hardware: 2-core CPU, no GPU; OMP_NUM_THREADS=2
• Seed: 0
• See args.yaml for full args.

Dataset

CC-BY 4.0 dataset, downloaded at training time via scripts/download_dataset.py. Source images are NOT redistributed here.

Usage

import onnxruntime as ort
sess = ort.InferenceSession("best.onnx", providers=["CPUExecutionProvider"])
# input: (1, 3, 320, 320) float32 RGB, normalised to [0,1]
# output: (1, N, 84) or YOLOv8 detection head — see Ultralytics docs.

Or via Ultralytics:

from ultralytics import YOLO
model = YOLO("best.onnx")
results = model("image.jpg", imgsz=320)

Caveat (round-1, fixed in round-2)

Vest class had 0 instances in the alphabetical 1200-image subset (all ds1_* files are helmet/head only). Round-2 will use class-balanced sampling.

License

MIT for these weights. Source dataset CC-BY 4.0 (see upstream TRAINING_REPORT.md for attribution). Ultralytics base model is AGPL-3.0 — using these weights in a hosted service may require AGPL compliance; see Ultralytics License.