Fine-Tune RF-DETR on Custom Classes from Open Images V7 dataset

Want to use the latest SOTA object detection models from Roboflow on custom classes? First, check if your classes exist in the Open Images dataset. If they do, good news! This repo shows you how to fine-tune a RF-DETR model with a single command.

Feel free to share any feedback, bugs, or improvement ideas!

Requirements

• uv (Python package and project manager). If not, just run (on macOS and Linux):

curl -LsSf https://astral.sh/uv/install.sh | sh

• Python 3.10+. If not, just run:

uv python install 3.10.18

• Download the downloader file

wget https://raw.githubusercontent.com/openimages/dataset/master/downloader.py

Check Astral documentation if you need alternative installation methods.

Usage

Quickstart

First, select your target class(es) from the 601 available in Open Images V7. View the set of boxable classes as a hierarchy here, or explore the dataset visually using the Open Images Bounding Boxes Explorer.

Clone the repository and navigate into it:

git clone --depth 1 https://github.com/PierreMarieCurie/train-rf-detr-OIDv7.git
cd train-rf-detr-OIDv7/

For example, to fine-tune RF-DETR on Tiger and Cheetah classes from Open Images V7:

uv run train.py \
  --target-classes tiger cheetah \
  --model base \
  --epochs 10 \
  --batch-size 8

This will:

• Download relevant annotations and images
• Convert them to COCO format
• Fine-tune RF-DETR on the selected subset
• Save logs and checkpoints

If you restart a training session, whether with the same or different classes, annotation files and already downloaded images will not be downloaded again.

Using the fine-tuned model

Once your RF-DETR model is fine-tuned, you can run inference on a new image using the infer.py script:

uv run infer.py \
  --image path/to/my/image.png \
  --model base \
  --checkpoint path/to/my/checkoint.pth \
  --output output.png

This script takes a trained checkpoint, runs detection on a single image, and saves the annotated result.

Advanced training settings

You can customize training with the following arguments:

Argument	Description
--target-classes	(Required) One or more class names to fine-tune on (e.g., tiger cheetah)
--model	Type of RFDETR model. Can be nano, small, medium, base or large (default: base)
--epochs	Number of training epochs (default: 20)
--learning-rate / --lr	Learning rate (default: 1e-4)
--batch-size	Batch size. (default: 20)
--grad-accum-steps	Gradient accumulation steps (default: 1)
--early-stopping	Early stepping (Add --early-stopping to enable it)

Other parameters or files will be downloaded or generated:

Argument	Description
--result-folder	Custom path to save results. If not set, a timestamped folder with class names is auto-created in results folder
--dataset-folder	Folder where the COCO-converted dataset will be stored (default: dataset)
--csv-folder	Folder where CSVs from Open Images will be saved (default: OIDv7_csv)
--manifest-path	Path to the Open Images V7 manifest JSON with links to the annotation files (default: csv_manifest.txt)

Alternatively, run:

uv run train.py -h

Advanced inference settings

You can customize training with the following arguments:

Argument	Description
--image	(Required) Path to the input image
--model	(Required) Type of RFDETR model. Can be nano, small, medium, base or large
--checkpoint	(Required) Path to checkpoint .pth file
--output	Path to save annotated output image" (default: detections.png)

Alternatively, run:

uv run infer.py -h

TODO

• Publish a Docker image to containerize the training phase

License

This repo is licensed under a MIT License.

Please note that OIDv7 annotations are licensed by Google LLC under CC BY 4.0 license. The images are listed as having a CC BY 2.0 license.

Also, RF-DETR is licensed under a Apache 2.0 license.

Acknowledgments

• This project builds on RF-DETR, a state-of-the-art, open-source object detection model released by Roboflow.

• This project uses uv by Astral.