Model Framework Comparison

License Comparison

Framework	License	Commercial Use	OAK-D Support
RT-DETR	Apache 2.0	✅ Free	⭐ Excellent
YOLOv6	MIT	✅ Free	⭐ Excellent
YOLOX	MIT	✅ Free	⭐ Excellent
RF-DETR	Check repo	⚠️ Unknown	⚠️ May need conversion
YOLOv8/v11	AGPL-3.0	❌ Paid ($1k-5k/yr)	Excellent

Performance on OAK-D 4 Pro (48 TOPS INT8)

Model	Size	Speed (FPS)	Accuracy	Training Time
RT-DETR r18	~15MB	30-40	Good	Fast
RT-DETR r34	~30MB	20-30	Better	Medium
YOLOv6n	~10MB	40-50	Good	Fast
YOLOv6s	~20MB	30-40	Better	Medium
YOLOX nano	~6MB	50-60	Good	Fast
YOLOX-s	~18MB	35-45	Better	Medium

Which to Choose?

For Maximum Speed (50-60 FPS):

YOLOX nano - Smallest, fastest, proven

For Best Balance (30-40 FPS):

RT-DETR r18 or YOLOv6n - Modern, accurate

For Best Accuracy (20-30 FPS):

RT-DETR r34 or YOLOv6s - Larger models

Recommended Starting Point:

YOLOv6n - Great balance, proven OAK compatibility, MIT license

Training Commands

All models use the same workflow in the GUI, or from command line:

RT-DETR

.venv/bin/python train_rtdetr.py \
    --dataset-dir dataset_prepared \
    --model rtdetr-r18 \
    --epochs 100

YOLOv6

.venv/bin/python train_yolov6.py \
    --dataset-dir dataset_prepared \
    --model yolov6n \
    --epochs 100

YOLOX (YOLOv8 equivalent)

.venv/bin/python train_yolox.py \
    --dataset-dir dataset_prepared \
    --model yolox-nano \
    --epochs 100

Export for OAK-D

All models export to OpenVINO format for OAK deployment:

# RT-DETR
.venv/bin/python export_rtdetr_oak.py --weights runs/rtdetr_training/training/weights/best.pt

# YOLOv6/YOLOX use Ultralytics export
.venv/bin/python -c "
from ultralytics import YOLO
model = YOLO('runs/yolov6_training/training/weights/best.pt')
model.export(format='openvino', imgsz=640, half=False)
"

Then convert to blob:

Online: https://blobconverter.luxonis.com/
CLI: blobconverter --openvino-xml model.xml

Tips

Start with nano/r18 models for fast iteration
Train for 100-200 epochs - use early stopping
Collect 200+ images for good accuracy
Test on OAK-D before collecting more data
Use INT8 quantization for full 48 TOPS speed

2.5 KiB Raw Blame History