DETR models in ONNX format

Model Zoo

Algorithm	Supported	Year	Github Link	Note
DETR	✅	2020	Link
Deformable-DETR	✅	2021	Link
LW-DETR	✅	2024	Link
RT-DETR	✅	2024	Link	Use focal loss
RF-DETR	✅	2025	Link
DEIMv2	✅	2025	Link	Use focal loss

Download weights from Release, then put them into models folder.

NOTE: The ONNX models already fused the image normalization using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].

Model info

The input shape of these models are dynamic (batch size and image size).

• Inputs:
  • input: ['N', 3, 'H', 'W'] (FLOAT)
• Outputs:
  • logits: ['N', 300, 80 or 91 or 301] (FLOAT)
  • boxes: ['N', 300, 4] (FLOAT)

Some models use focal loss instead of cross entropy loss.

You can view ONNX models with netron, which can visualize the shape of each layer.

With onnxruntime, you can use the following script to get some model infos.

import onnxruntime as ort

sess = ort.InferenceSession(model_path)
input_name = sess.get_inputs()[0].name
output_names = [output.name for output in sess.get_outputs()]
input_shape = sess.get_inputs()[0].shape
meta = sess.get_modelmeta().custom_metadata_map
stride = int(meta.get('stride', -1))
class_names = eval(meta.get('names', '{}'))

Usage

python

import cv2 as cv
from detr_onnx import DetrONNX

model_path = "models/rt-detrv2.onnx"
detr = DetrONNX(model_path)

img = cv.imread("images/kite.jpg")
prob, boxes = detr.detect(img)
res = detr.plot_result(img, prob, boxes)

cv.imshow("res", res)
cv.waitKey(0)

C++

mkdir build && cd build
cmake ../cpp
# Usage: ./build/detr_onnx <model_path> <image_path>