Improving YOLO Model for Four-Leaf Clover Detection

Been working on a four-leaf clover detection model for an iOS app. Had a working YOLOv8n model but it was stuck at 58% mAP@0.5. Not great when you’re trying to spot tiny clovers in grass.

The problem

Current model (YOLOv8n) trained for 50 epochs and early-stopped:

• mAP@0.5: 58.4%
• Precision: 80.2%
• Recall: 52.6%

Missing half the clovers isn’t going to cut it.

What I changed

1. Switched to YOLOv8s-P2

The -P2 variant adds an extra detection head at stride 4 (instead of starting at stride 8). This is specifically designed for small object detection.

model = YOLO("yolov8s-p2.yaml").load("yolov8s.pt")

2. Fixed early stopping

Training was stopping too early. Bumped patience from 20 to 50:

epochs=200,
patience=50,

3. Learning rate schedule

The default lrf=0.01 drops learning rate too aggressively. Changed to cosine annealing with gentler decay:

lrf=0.1,
cos_lr=True,
warmup_epochs=5.0,

4. Augmentation boost

Clovers look the same upside down, so vertical flip is free data:

flipud=0.5,      # was 0.0
degrees=25.0,    # was 15.0
mixup=0.15,      # was 0.0
copy_paste=0.1,  # was 0.0
multi_scale=True,

The copy_paste augmentation is nice for small datasets - it copies annotated objects onto other images.

Full training config

from ultralytics import YOLO

model = YOLO("yolov8s-p2.yaml").load("yolov8s.pt")

results = model.train(
    data="dataset.yaml",
    epochs=200,
    patience=50,
    imgsz=640,
    batch=8,
    device=0,  # GPU

    lr0=0.01,
    lrf=0.1,
    cos_lr=True,
    warmup_epochs=5.0,
    cls=0.75,

    flipud=0.5,
    degrees=25.0,
    scale=0.7,
    translate=0.2,
    hsv_h=0.02,
    hsv_v=0.5,

    mosaic=1.0,
    close_mosaic=15,
    mixup=0.15,
    copy_paste=0.1,
    multi_scale=True,

    optimizer="AdamW",
    project="runs",
    name="yolov8s_p2_improved",
)

Also trying 768 resolution

Created a second experiment with imgsz=768 instead of 640. Higher resolution should help with small objects, but takes longer to train. Running both to compare.

Key difference for 768 version:

imgsz=768,
batch=12,  # reduced due to memory

What’s next

If neither hits 80% mAP, backup plans:

1. SAHI - slice images into tiles for inference
2. Pseudo-labeling - use model to label the 294 unlabeled images, retrain
3. Hard negative mining - collect false positives (3-leaf clovers) and add as negatives

Training on Colab with T4 GPU. Will update with results.