rtdetrv2_pytorch

Quick start

Setup

pip install -r requirements.txt

The following is the corresponding torch and torchvision versions.

rtdetr	torch	torchvision
-	2.4	0.19
-	2.2	0.17
-	2.1	0.16
-	2.0	0.15

Fig

Model Zoo

Base models

Model	Dataset	Input Size	APval	AP50val	#Params(M)	FPS	config	checkpoint
RT-DETRv2-S	COCO	640	48.1 (+1.6)	65.1	20	217	config	url
RT-DETRv2-M*	COCO	640	49.9 (+1.0)	67.5	31	161	config	url
RT-DETRv2-M	COCO	640	51.9 (+0.6)	69.9	36	145	config	url
RT-DETRv2-L	COCO	640	53.4 (+0.3)	71.6	42	108	config	url
RT-DETRv2-X	COCO	640	54.3	72.8 (+0.1)	76	74	config	url

Notes:

• AP is evaluated on MSCOCO val2017 dataset.
• FPS is evaluated on a single T4 GPU with $batch\_size = 1$, $fp16$, and $TensorRT>=8.5.1$.
• COCO + Objects365 in the table means finetuned model on COCO using pretrained weights trained on Objects365.

Models of discrete sampling

Model	Sampling Method	APval	AP50val	config	checkpoint
RT-DETRv2-S_dsp	discrete_sampling	47.4	64.8 (-0.1)	config	url
RT-DETRv2-M*_dsp	discrete_sampling	49.2	67.1 (-0.4)	config	url
RT-DETRv2-M_dsp	discrete_sampling	51.4	69.7 (-0.2)	config	url
RT-DETRv2-L_dsp	discrete_sampling	52.9	71.3 (-0.3)	config	url

Notes:

• The impact on inference speed is related to specific device and software.
• *_dsp* is the model inherit *_sp* model's knowledge and adapt to discrete_sampling strategy. You can use TensorRT 8.4 (or even older versions) to inference for these models

Ablation on sampling points

Model	Sampling Method	#Points	APval	AP50val	checkpoint
rtdetrv2_r18vd_sp1	grid_sampling	21,600	47.3	64.3 (-0.6)	url
rtdetrv2_r18vd_sp2	grid_sampling	43,200	47.7	64.7 (-0.2)	url
rtdetrv2_r18vd_sp3	grid_sampling	64,800	47.8	64.8 (-0.1)	url
rtdetrv2_r18vd(_sp4)	grid_sampling	86,400	47.9	64.9	url

Notes:

• The impact on inference speed is related to specific device and software.
• #points the total number of sampling points in decoder for per image inference.

Usage

details

1. Training

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=9909 --nproc_per_node=4 tools/train.py -c path/to/config --use-amp --seed=0 &> log.txt 2>&1 &

2. Testing

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=9909 --nproc_per_node=4 tools/train.py -c path/to/config -r path/to/checkpoint --test-only

3. Tuning

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=9909 --nproc_per_node=4 tools/train.py -c path/to/config -t path/to/checkpoint --use-amp --seed=0 &> log.txt 2>&1 &

4. Export onnx

python tools/export_onnx.py -c path/to/config -r path/to/checkpoint --check

5. Inference

Support torch, onnxruntime, tensorrt and openvino, see details in references/deploy

python references/deploy/rtdetrv2_onnx.py --onnx-file=model.onnx --im-file=xxxx
python references/deploy/rtdetrv2_tensorrt.py --trt-file=model.trt --im-file=xxxx
python references/deploy/rtdetrv2_torch.py -c path/to/config -r path/to/checkpoint --im-file=xxx --device=cuda:0

Citation

If you use RTDETR or RTDETRv2 in your work, please use the following BibTeX entries:

bibtex

@misc{lv2023detrs,
      title={DETRs Beat YOLOs on Real-time Object Detection},
      author={Wenyu Lv and Shangliang Xu and Yian Zhao and Guanzhong Wang and Jinman Wei and Cheng Cui and Yuning Du and Qingqing Dang and Yi Liu},
      year={2023},
      eprint={2304.08069},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{lv2024rtdetrv2improvedbaselinebagoffreebies,
      title={RT-DETRv2: Improved Baseline with Bag-of-Freebies for Real-Time Detection Transformer}, 
      author={Wenyu Lv and Yian Zhao and Qinyao Chang and Kui Huang and Guanzhong Wang and Yi Liu},
      year={2024},
      eprint={2407.17140},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.17140}, 
}