YOLO-DC: Enhancing object detection with deformable convolutions and contextual mechanism

Introduction

YOLO-DC outperforms numerous state-of-the-art (SOTA) algorithms, including YOLOv8, while maintaining a comparable level of computation and parameter count. For more details, please refer to our report on Github. The related paper has been accepted for publication in Signal Processing: Image Communication.

In the figure above, (a) and (b) depict comparisons of computational and parameter counts among the models on the COCO 2017 dataset, respectively.

Benchmark

|Model |size |AP^{val
0.5:0.95 |APval
0.5 | Params
(M) |FLOPs
(G)| | ------ |:---: | :---: | :---: |:---: | :---: | |YOLO-DC-N |640 |40.8 |56.9 |3.9 | 8.9 | |YOLO-DC-S |640 |46.6 |63.5 |13.9 | 29.2 | |YOLO-DC-M |640 |50.4 |67.3 |32.9 | 70.9 |}

Table Notes

• Results of the AP and speed are evaluated on COCO val2017 dataset with the input resolution of 640×640.
• All experiments are based on NVIDIA 3090 GPU.

Environment

• python requirements

shell pip install -r requirements.txt If you are prompted that a package is missing, follow the corresponding prompts to follow that package. - data:

prepare COCO dataset, YOLO format coco labels and specify dataset paths in data.yaml (data.yaml is located at ". /ultralytics/datasets/coco.yaml").

Train

### 1. command-line mode - See train.py for more information on how to use it.

shell python ./train.py --yaml ./ultralytics/models/v8/YOLO-DC.yaml --data ./ultralytics/datasets/coco.yaml --weight path --epoch 500 --device 0,1,2,3,4 --batch 128 ### 2. python Writing direct Python code such as main.py is an example. ```shell from ultralytics import YOLO def main(): # 加载模型 model = YOLO("./ultralytics/models/v8/yolov8n-DC.yaml") # 从头开始构建新模型 # model = YOLO("./runs/detect/train_500/weights/best.pt") # 加载预训练模型（建议用于训练）

# 使用模型
model.train(data="./ultralytics/datasets/coco.yaml",
            epochs=500, device='cuda:0',
            batch=48,
            save_period=50,
            verbose=True,
            project="COCO",
            name="train_DC_n_500",
            profile=True,)  # 训练模型
metrics = model.val(name="val_DC_n_500")  # 在验证集上评估模型性能

if name == 'main': main()

```

Cite

If you feel our work is helpful to you, please cite the following paper: ``` @INPROCEEDINGS{YOLO-DC, author={Zhang, Dengyong and Xu, Chuanzhen and Chen, Jiaxin and Deng, Bin and Liao, Xin}, booktitle={2024 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)}, title={YOLO-DC: Enhancing object detection with deformable convolutions and contextual mechanism}, year={2024}, volume={}, number={}, pages={1-6}, keywords={Convolutional codes;Deformable models;Convolution;Source coding;Data preprocessing;Object detection;Information processing;Feature extraction;Data models;Detection algorithms}, doi={10.1109/APSIPAASC63619.2025.10848905}}

@article{ZHANG2025117373, title = {YOLO-DC: Integrating deformable convolution and contextual fusion for high-performance object detection}, journal = {Signal Processing: Image Communication}, pages = {117373}, year = {2025}, issn = {0923-5965}, doi = {https://doi.org/10.1016/j.image.2025.117373}, url = {https://www.sciencedirect.com/science/article/pii/S0923596525001195}, author = {Dengyong Zhang and Chuanzhen Xu and Jiaxin Chen and Lei Wang and Bin Deng}, keywords = {Object detection, YOLO, Deformable convolutions, Contextual mechanisms, Deep learning} } ```

Acknowledgement

The implementation is based on YOLOv8. Thanks for their open source code.