DEIM

---

🚀 Updates

• [2025.09.26] DEIMv2 is now available with the project page and release code. The series covers eight model sizes, from X down to Atto. For the S, M, L, and X variants, we leverage DINOv3 features (distilled or pretrained). DEIMv2 achieves higher performance with fewer parameters and FLOPs.
• [2025.06.24] DEIMv2 is coming soon: our next-gen detection series, along with three ultra-light variants: Pico (1.5M), Femto (0.96M), and Atto (0.49M), all delivering SoTA performance. Atto, in particular, is tailored for mobile devices, achieving 23.8 AP on COCO at 320×320 resolution.
• [2025.03.12] The Object365 Pretrained DEIM-D-FINE-X model is released, which achieves 59.5% AP after fine-tuning 24 COCO epochs.
• [2025.03.05] The Nano DEIM model is released.
• [2025.02.27] The DEIM paper is accepted to CVPR 2025. Thanks to all co-authors.
• [2024.12.26] A more efficient implementation of Dense O2O, achieving nearly a 30% improvement in loading speed (See the pull request for more details). Huge thanks to my colleague Longfei Liu.
• [2024.12.03] Release DEIM series. Besides, this repo also supports the re-implmentations of D-FINE and RT-DETR.

Table of Content

• 1. Model Zoo
• 2. Quick start
• 3. Usage
• 4. Tools
• 5. Citation
• 6. Acknowledgement

1. Model Zoo

DEIM-D-FINE

Model	Dataset	APD-FINE	APDEIM	#Params	Latency	GFLOPs	config	checkpoint
N	COCO	42.8	43.0	4M	2.12ms	7	yml	ckpt
S	COCO	48.7	49.0	10M	3.49ms	25	yml	ckpt
M	COCO	52.3	52.7	19M	5.62ms	57	yml	ckpt
L	COCO	54.0	54.7	31M	8.07ms	91	yml	ckpt
X	COCO	55.8	56.5	62M	12.89ms	202	yml	ckpt

DEIM-RT-DETRv2

Model	Dataset	APRT-DETRv2	APDEIM	#Params	Latency	GFLOPs	config	checkpoint
S	COCO	47.9	49.0	20M	4.59ms	60	yml	ckpt
M	COCO	49.9	50.9	31M	6.40ms	92	yml	ckpt
M*	COCO	51.9	53.2	33M	6.90ms	100	yml	ckpt
L	COCO	53.4	54.3	42M	9.15ms	136	yml	ckpt
X	COCO	54.3	55.5	76M	13.66ms	259	yml	ckpt

2. Quick start

Setup

conda create -n deim python=3.11.9
conda activate deim
pip install -r requirements.txt

Data Preparation

1. Download COCO2017 from OpenDataLab or COCO.

2. Modify paths in coco_detection.yml

   train_dataloader:
       img_folder: /data/COCO2017/train2017/
       ann_file: /data/COCO2017/annotations/instances_train2017.json
   val_dataloader:
       img_folder: /data/COCO2017/val2017/
       ann_file: /data/COCO2017/annotations/instances_val2017.json
   

To train on your custom dataset, you need to organize it in the COCO format. Follow the steps below to prepare your dataset:

1. Set remap_mscoco_category to False:

   This prevents the automatic remapping of category IDs to match the MSCOCO categories.

   remap_mscoco_category: False
   

2. Organize Images:

   Structure your dataset directories as follows:

   dataset/
   ├── images/
   │   ├── train/
   │   │   ├── image1.jpg
   │   │   ├── image2.jpg
   │   │   └── ...
   │   ├── val/
   │   │   ├── image1.jpg
   │   │   ├── image2.jpg
   │   │   └── ...
   └── annotations/
       ├── instances_train.json
       ├── instances_val.json
       └── ...
   

   • images/train/: Contains all training images.
   • images/val/: Contains all validation images.
   • annotations/: Contains COCO-formatted annotation files.

3. Convert Annotations to COCO Format:

   If your annotations are not already in COCO format, you'll need to convert them. You can use the following Python script as a reference or utilize existing tools:

   import json
   
   def convert_to_coco(input_annotations, output_annotations):
       # Implement conversion logic here
       pass
   
   if __name__ == "__main__":
       convert_to_coco('path/to/your_annotations.json', 'dataset/annotations/instances_train.json')
   

4. Update Configuration Files:

   Modify your custom_detection.yml.

   task: detection
   
   evaluator:
     type: CocoEvaluator
     iou_types: ['bbox', ]
   
   num_classes: 777 # your dataset classes
   remap_mscoco_category: False
   
   train_dataloader:
     type: DataLoader
     dataset:
       type: CocoDetection
       img_folder: /data/yourdataset/train
       ann_file: /data/yourdataset/train/train.json
       return_masks: False
       transforms:
         type: Compose
         ops: ~
     shuffle: True
     num_workers: 4
     drop_last: True
     collate_fn:
       type: BatchImageCollateFunction
   
   val_dataloader:
     type: DataLoader
     dataset:
       type: CocoDetection
       img_folder: /data/yourdataset/val
       ann_file: /data/yourdataset/val/ann.json
       return_masks: False
       transforms:
         type: Compose
         ops: ~
     shuffle: False
     num_workers: 4
     drop_last: False
     collate_fn:
       type: BatchImageCollateFunction
   

3. Usage

1. Training

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=7777 --nproc_per_node=4 train.py -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml --use-amp --seed=0

2. Testing

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=7777 --nproc_per_node=4 train.py -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml --test-only -r model.pth

3. Tuning

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=7777 --nproc_per_node=4 train.py -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml --use-amp --seed=0 -t model.pth

For example, if you want to double the total batch size when training D-FINE-L on COCO2017, here are the steps you should follow:

1. Modify your dataloader.yml to increase the total_batch_size:

   train_dataloader:
       total_batch_size: 64  # Previously it was 32, now doubled
   

2. Modify your deim_hgnetv2_l_coco.yml. Here’s how the key parameters should be adjusted:

   optimizer:
   type: AdamW
   params:
       -
       params: '^(?=.*backbone)(?!.*norm|bn).*$'
       lr: 0.000025  # doubled, linear scaling law
       -
       params: '^(?=.*(?:encoder|decoder))(?=.*(?:norm|bn)).*$'
       weight_decay: 0.
   
   lr: 0.0005  # doubled, linear scaling law
   betas: [0.9, 0.999]
   weight_decay: 0.0001  # need a grid search
   
   ema:  # added EMA settings
       decay: 0.9998  # adjusted by 1 - (1 - decay) * 2
       warmups: 500  # halved
   
   lr_warmup_scheduler:
       warmup_duration: 250  # halved
   

If you'd like to train DEIM on COCO2017 with an input size of 320x320, follow these steps:

1. Modify your dataloader.yml:

   
   train_dataloader:
   dataset:
       transforms:
           ops:
               - {type: Resize, size: [320, 320], }
   collate_fn:
       base_size: 320
   dataset:
       transforms:
           ops:
               - {type: Resize, size: [320, 320], }
   

2. Modify your dfine_hgnetv2.yml:

   eval_spatial_size: [320, 320]
   

4. Tools

1. Setup

pip install onnx onnxsim

2. Export onnx

python tools/deployment/export_onnx.py --check -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml -r model.pth

3. Export tensorrt

trtexec --onnx="model.onnx" --saveEngine="model.engine" --fp16

1. Setup

pip install -r tools/inference/requirements.txt

2. Inference (onnxruntime / tensorrt / torch)

Inference on images and videos is now supported.

python tools/inference/onnx_inf.py --onnx model.onnx --input image.jpg  # video.mp4
python tools/inference/trt_inf.py --trt model.engine --input image.jpg
python tools/inference/torch_inf.py -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml -r model.pth --input image.jpg --device cuda:0

1. Setup

pip install -r tools/benchmark/requirements.txt

2. Model FLOPs, MACs, and Params

python tools/benchmark/get_info.py -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml

2. TensorRT Latency

python tools/benchmark/trt_benchmark.py --COCO_dir path/to/COCO2017 --engine_dir model.engine

1. Setup

pip install fiftyone

4. Voxel51 Fiftyone Visualization (fiftyone)

python tools/visualization/fiftyone_vis.py -c configs/deim_dfine/deim_hgnetv2_${model}_coco.yml -r model.pth

1. Auto Resume Training

bash reference/safe_training.sh

2. Converting Model Weights

python reference/convert_weight.py model.pth

5. Citation

If you use DEIM or its methods in your work, please cite the following BibTeX entries:

@misc{huang2024deim,
      title={DEIM: DETR with Improved Matching for Fast Convergence},
      author={Shihua, Huang and Zhichao, Lu and Xiaodong, Cun and Yongjun, Yu and Xiao, Zhou and Xi, Shen},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
      year={2025},
}

6. Acknowledgement

Our work is built upon D-FINE and RT-DETR.

✨ Feel free to contribute and reach out if you have any questions! ✨