opencv

/

opencv_contribution

like 3

Follow

Open Source Vision Foundation 149

Add Ultralytics YOLO models in TFLite format

Adds four Ultralytics YOLO models exported to TFLite, runnable through OpenCV's DNN TFLite importer

Models

Each folder contains

• <model>.tflite — the exported model (fp32).
• demo.py — inference using OpenCV DNN only (cv2.dnn.readNet): letterbox preprocessing matching
  ultralytics.data.augment.LetterBox, head decode (classic or NMS-free, including mask reconstruction for -seg),
  NMS, and an annotated output image.
• README.md — the .pt → .tflite conversion recipe and usage.
• LICENSE — AGPL-3.0 (Ultralytics).
• example_outputs/ — input image + annotated result.

Conversion

Each model was exported from its Ultralytics .pt checkpoint with:

yolo export model=<model>.pt format=tflite imgsz=640

or the Python API:

from ultralytics import YOLO
YOLO("<model>.pt").export(format="tflite", imgsz=640)

Replace <model> with yolov8n, yolov5nu, yolo26n, or yolo26n-seg. See each folder's README.md for the exact
command and the pinned environment/versions used to produce that model.

Usage

python demo.py --model <model>.tflite --image example_outputs/input.jpg --output example_outputs/<model>_output.jpg

License

Weights and the Ultralytics exporter are released under AGPL-3.0 — see each folder's LICENSE and the
Ultralytics LICENSE.