femartip/yolo-zod

Model Card: YOLOv11s on ZOD (Vehicle, VulnerableVehicle, Pedestrian)

Model Details

• Architecture: YOLOv11l (Ultralytics), initialized from pretrained weights
• Framework: Ultralytics YOLO

Training Data

• Source: Zenseact Open Dataset (ZOD), converted to COCO, then to YOLO
• Classes kept: Vehicle, VulnerableVehicle, Pedestrian (class order in data/zod_yolo/dataset.yaml)
• Filtering: exclude occlusion levels Heavy and VeryHeavy (keep None/Medium/Light); drop boxes with height < 25 px
• Splits: original ZOD train is split into train/val by 1 km lat/lon tiles (80/20; 69,090 train / 20,882 val) to reduce spatial leakage and keep weather/road_type/time_of_day distributions similar; original ZOD val -> YOLO test

Training Procedure

• Key settings from args.yaml:
  • Image size 512, batch 32, epochs 200 (early stopped at epoch 154 with patience )
  • Base LR 0.1, seed 43

Evaluation Results

Metrics (epoch 262, on my defined val split):

• AP50: 46.6
• AP50-95: 38.4

Usage (Ultralytics YOLO)

from ultralytics import YOLO

model = YOLO("final_model.pt")
results = model.predict(
    source="path/to/image.jpg",
    imgsz=512,
    conf=0.25,
)

# Class names (matches data/zod_yolo/dataset.yaml)
names = {0: "Vehicle", 1: "VulnerableVehicle", 2: "Pedestrian"}

Limitations

• Trained on only three classes; other ZOD object types are not represented.
• Heavy/VeryHeavy occlusions and very small objects (<25 px height) are filtered out and may perform poorly at inference time.