I’m training YOLOv8n on a custom dataset (5000 images, 3 classes) using Ultralytics. Despite high confidence scores (0.8+), the model misses obvious objects in validation. Dataset Structure dataset/ ├── images/ └── labels/ Training Command: model.train(data="data.yaml", epochs=100, batch=16, imgsz=640) Issue: mAP@0.5 is only 0.35 (expected ≥0.7). Validation images show missed detections even when confidence is high. Tried: Reduced lr0 from 0.01 to 0.001. Verified annotations in LabelImg (no errors found). Added mosaic augmentation (mosaic=1.0). Question: What could cause high confidence but poor detection performance? Are there key hyperparameters I should adjust? Continue reading...
