RF-DETR: An Open Source Model for Real-Time Visual Object Detection

RF-DETR is an open source object detection model developed by the Roboflow team. It is based on Transformer architecture, the core feature is real-time efficiency. The model achieves real-time detection over 60 APs for the first time on the Microsoft COCO dataset, and also performs outstandingly in the RF100-VL benchmark test, adapting to a variety of real-world scenarios. It is available in two versions: RF-DETR-base (29 million parameters) and RF-DETR-large (128 million parameters). The model is small and suitable for edge device deployment. The code and pre-trained weights are licensed under the Apache 2.0 license and are free and open for community use. Users can obtain resources from GitHub for easy training or deployment.

Function List

• Real-time object detection: fast recognition of objects in images or videos with low latency.
• Custom dataset training: support for tuning models with your own data.
• Running on edge devices: the model is lightweight and suitable for resource-limited devices.
• Adjustable resolution: users can balance inspection speed and accuracy.
• Pre-training model support: provides pre-trained weights based on the COCO dataset.
• Video stream processing: can analyze the video in real time and output the results.
• ONNX Export: Supports conversion to ONNX format for easy cross-platform deployment.
• Multi-GPU training: You can accelerate the training process with multiple graphics cards.

Using Help

The use of RF-DETR is divided into three parts: installation, inference and training. Below are detailed steps to help you get started quickly.

Installation process

1. environmental preparation
   Requires Python 3.9 or higher, and PyTorch 1.13.0 or higher. If using a GPU, run nvidia-smi Check the drive.
   • Install PyTorch:

     pip install torch>=1.13.0 torchvision>=0.14.0
     

   • Download code:

     git clone https://github.com/roboflow/rf-detr.git
     cd rf-detr
     

   • Install the dependencies:

     pip install rfdetr
     

     This will automatically install numpy,supervision and other necessary libraries.

2. Verify Installation
   Run the following code:

   from rfdetr import RFDETRBase
   print("安装成功")

If no errors are reported, the installation is complete.

inference operation

RF-DETR comes with a pre-trained model of the COCO dataset to detect images or videos directly.

1. image detection
   • Sample code:

     import io
     import requests
     from PIL import Image
     from rfdetr import RFDETRBase
     import supervision as sv
     model = RFDETRBase()
     url = "https://media.roboflow.com/notebooks/examples/dog-2.jpeg"
     image = Image.open(io.BytesIO(requests.get(url).content))
     detections = model.predict(image, threshold=0.5)
     labels = [f"{class_id} {confidence:.2f}" for class_id, confidence in zip(detections.class_id, detections.confidence)]
     annotated_image = image.copy()
     annotated_image = sv.BoxAnnotator().annotate(annotated_image, detections)
     annotated_image = sv.LabelAnnotator().annotate(annotated_image, detections, labels)
     sv.plot_image(annotated_image)
     

   • This code detects objects in the image, labels the bounding box and confidence level, and then displays the results.
2. Video Detection
   • first install opencv-python::

     pip install opencv-python
     

   • Sample code:

     import cv2
     from rfdetr import RFDETRBase
     import supervision as sv
     model = RFDETRBase()
     cap = cv2.VideoCapture("video.mp4")  # 替换为你的视频路径
     while cap.isOpened():
     ret, frame = cap.read()
     if not ret:
     break
     image = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
     detections = model.predict(image, threshold=0.5)
     annotated_frame = sv.BoxAnnotator().annotate(frame, detections)
     cv2.imshow("RF-DETR Detection", annotated_frame)
     if cv2.waitKey(1) & 0xFF == ord('q'):
     break
     cap.release()
     cv2.destroyAllWindows()
     

   • This will detect objects in the video frame by frame and display them in real time.
3. Adjustment of resolution
   • The resolution can be set at initialization (must be a multiple of 56):

     model = RFDETRBase(resolution=560)
     

   • The higher the resolution, the better the accuracy, but it will be slower.

Training customized models

RF-DETR supports fine-tuning with its own dataset, but the dataset needs to be in COCO format, containing train,valid cap (a poem) test Three subdirectories.

1. Preparing the dataset
   • Example catalog structure:

     dataset/
     ├── train/
     │   ├── _annotations.coco.json
     │   ├── image1.jpg
     │   └── image2.jpg
     ├── valid/
     │   ├── _annotations.coco.json
     │   ├── image1.jpg
     │   └── image2.jpg
     └── test/
     ├── _annotations.coco.json
     ├── image1.jpg
     └── image2.jpg
     

   • COCO format datasets can be generated using the Roboflow platform:

     from roboflow import Roboflow
     rf = Roboflow(api_key="你的API密钥")
     project = rf.workspace("rf-100-vl").project("mahjong-vtacs-mexax-m4vyu-sjtd")
     dataset = project.version(2).download("coco")
     

2. Start training
   • Sample code:

     from rfdetr import RFDETRBase
     model = RFDETRBase()
     model.train(dataset_dir="./mahjong-vtacs-mexax-m4vyu-sjtd-2", epochs=10, batch_size=4, grad_accum_steps=4, lr=1e-4)
     

   • For training, the recommended total batch size (batch_size * grad_accum_stepsFor example, the A100 GPUs use the batch_size=16, grad_accum_steps=1T4 GPUs batch_size=4, grad_accum_steps=4The
3. Multi-GPU Training
   • establish main.py Documentation:

     from rfdetr import RFDETRBase
     model = RFDETRBase()
     model.train(dataset_dir="./dataset", epochs=10, batch_size=4, grad_accum_steps=4, lr=1e-4)
     

   • Runs in the terminal:

     python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py
     

   • commander-in-chief (military) 8 Replace with the number of GPUs you are using. Note the adjustment batch_size to keep the total batch size stable.
4. Load training results
   • Two weight files are generated after training: regular weights and EMA weights (more stable). Loading method:

     model = RFDETRBase(pretrain_weights="./output/model_ema.pt")
     detections = model.predict("image.jpg")
     

ONNX Export

• Export to ONNX format for easy deployment on other platforms:

  from rfdetr import RFDETRBase
  model = RFDETRBase()
  model.export()
  

• The exported file is saved in the output Directory for optimized reasoning for edge devices.

application scenario

1. automatic driving
   RF-DETR detects vehicles and pedestrians on the road in real time. Its low latency and high accuracy are suitable for embedded systems.
2. industrial quality control
   RF-DETR quickly identifies part defects on factory assembly lines. The model is lightweight and can be run directly on the equipment.
3. video surveillance
   RF-DETR processes surveillance video to detect abnormal objects or behavior in real time. It supports video streaming and is suitable for 24/7 security.

QA

1. What dataset formats are supported?
   Only the COCO format is supported. The dataset needs to contain train,valid cap (a poem) test subdirectories, each with a corresponding _annotations.coco.json Documentation.
2. How to get Roboflow API key?
   Log in to https://app.roboflow.com, find the API key in your account settings, copy it and set it to the environment variable ROBOFLOW_API_KEYThe
3. How long does the training take?
   Depends on hardware and dataset size. On a T4 GPU, 10 epochs might take a couple hours. Smaller datasets can be run on a CPU, but it's slow.