🐝 Dataset Preparation: The Journey to Clean Data

The Dataset Challenge

Training an accurate object detection model requires high-quality, consistently labeled data. My journey began with four datasets from different sources, each with its own labeling conventions.

Dataset Merging Flow

The Label Conversion Bug

My first attempt to fix the V4 labels had a critical bug:

mapping = {'0': '1', '3': '0'}  # Wrong approach!
for line in lines:
    parts = line.strip().split()
    old_id = parts[0]
    if old_id in mapping:
        parts[0] = mapping[old_id]  # Bug: Sequential mapping causes collision

The Fix: Atomic Swap

for line in lines:
    parts = line.strip().split()
    if not parts:
        continue
    old_id = parts[0]
    if old_id == '0':
        parts[0] = '1'   # varroa (was 0) → 1
        modified = True
    elif old_id == '1':
        parts[0] = '0'   # bee (was 1) → 0
        modified = True

Class Distribution Analysis

Class Distribution

Image Size & IMGSZ Analysis

Understanding native image resolution is critical for choosing the right imgsz parameter. Upscaling low-resolution images creates artificial detail and slows training without improving accuracy.

Upscaling Impact by imgsz Setting

Varroa Bounding Box Analysis

Small objects (<16px) are notoriously difficult for YOLO to detect. This analysis determines how large varroa bounding boxes appear at different imgsz settings.

imgsz Decision Tree

Model & Training Setup

Model Selection: YOLOv8s

Baseline-First Training Strategy

Baseline Configuration

from ultralytics import YOLO

model = YOLO('yolov8s.pt')
results = model.train(
    data='varroa.yaml',
    epochs=100,
    imgsz=320,
    batch=16,
    patience=20,
    project='varroa_detection',
    name='baseline_v8s_imgsz320'
)

Key Takeaways