petrol-trained-v1

Fine-tuned YOLOv8s trained from scratch (COCO init) on 357 petrol-station images (yolo_high_confidence + filtered candidate_gt) for 120 epochs at 1280×1280 on an NVIDIA RTX A4000.

• Best val mAP@0.5 during training: 0.782 (epoch 114)
• Final val mAP@0.5: 0.7525, mAP@0.5–0.95: 0.5586
• 4 classes: petrol hose, petrol pump, price board, roof canopy
• ONNX FP16, 22.7 MB (≤ 30 MB rule)