Object Detection System Development

Object Detection System Development

Object detection is the task of simultaneous localization (bounding box) and classification of objects in an image. A single model in one pass produces: box coordinates, object class, confidence score. Applications include product counting on shelves, defect detection on conveyor lines, vehicle recognition, and human detection in video.

Detector Selection

YOLOv8/YOLO11 — optimal choice for most tasks. Ultralytics implementation with good documentation, active support, and built-in export to TensorRT/ONNX.

RT-DETR (Real-Time Detection Transformer) — transformer-based detector, better quality at comparable speed to YOLOv8.

Grounding DINO — open-vocabulary detection: finds objects by text description without retraining. Useful for prototyping and rare category tasks.

Fine-tuning for Custom Classes

from ultralytics import YOLO

# Load pretrained model
model = YOLO('yolov8l.pt')

# Train on custom dataset
results = model.train(
    data='dataset.yaml',      # dataset config path
    epochs=100,
    imgsz=640,
    batch=16,
    optimizer='AdamW',
    lr0=0.001,
    lrf=0.01,                 # final LR = lr0 * lrf
    weight_decay=0.0005,
    augment=True,
    degrees=10.0,             # rotation augmentation
    mosaic=1.0,               # mosaic augmentation
    device=0
)

dataset.yaml structure:

path: /data/myproject
train: images/train
val: images/val
test: images/test

nc: 5  # number of classes
names: ['cat', 'dog', 'car', 'person', 'bicycle']

Detection-Specific Augmentation

Detection requires specific augmentation — transformations must correctly apply to bounding boxes:

• Mosaic — combining 4 images into one, increases context diversity
• MixUp — blending two images with weights
• Copy-Paste — cutting objects and pasting in new context
• Random crop preserving objects in frame
• Albumentations: HorizontalFlip, RandomBrightnessContrast, GaussNoise

Detection Metrics

• [email protected] — mean Average Precision at IoU threshold 0.5
• [email protected]:0.95 — stricter: average mAP at IoU from 0.5 to 0.95 with 0.05 step
• Precision / Recall at specific confidence threshold
• FPS / latency — for real-time systems

Confidence threshold selection: ROC-like precision-recall curve, threshold choice depends on acceptable precision/recall balance for specific application.

NMS and Post-processing

Non-Maximum Suppression removes duplicate detections. Parameters: IoU threshold (0.45–0.7), confidence threshold (0.25–0.5). For densely located objects — Soft-NMS or Class-Agnostic NMS.

Deployment

TensorRT engine for NVIDIA GPU: export from Ultralytics with one command. ONNX for CPU deployment. For Raspberry Pi / Jetson: YOLO11n in TFLite / ONNX.