D-FINE

D-FINE redefines the regression task in DETRs as fine-grained distribution refinement.

CommonProductImageObject DetectionDETR

D-FINE is a powerful real-time object detection model that redefines the bounding box regression task in DETRs as fine-grained distribution refinement (FDR) and introduces Global Optimal Localization Self-Distillation (GO-LSD). It achieves outstanding performance without incurring additional inference and training costs. Developed by researchers from the Chinese Academy of Sciences, the model aims to enhance the accuracy and efficiency of object detection.

• Fine-grained Distribution Refinement (FDR): Achieves more accurate object localization through iterative refinement of probability distributions.
• Global Optimal Localization Self-Distillation (GO-LSD): Extracts localization knowledge from the refined distribution of the final layer and distills it to earlier layers using DDF loss and a decoupled weight strategy.
• Real-time Object Detection: D-FINE can perform real-time object detection while maintaining high accuracy.
• Model Series: Offers models of various sizes to accommodate different computational resources and latency requirements.
• Pre-trained Models: Provides models pre-trained on the COCO and Objects365 datasets
facilitating transfer learning.
• Open Source Code and Pre-trained Weights: Allows researchers and developers to freely use and modify.
• Support for Custom Dataset Training: Users can train models using custom datasets according to their needs.

The target audience for D-FINE includes researchers and developers in the field of computer vision
especially those focusing on object detection tasks. D-FINE is highly suitable for applications that require fast and accurate object localization
such as video surveillance
autonomous driving
and robotic vision

In video surveillance systems
D-FINE can be used for real-time detection and tracking of multiple targets.
In autonomous driving technology
D-FINE can identify and locate obstacles like pedestrians and vehicles on the road.
In robotic vision

1. Install the Python environment and necessary dependencies.
2. Clone the D-FINE repository to your local system.
3. Download pre-trained models as needed or train the models on a custom dataset.
4. Configure model parameters and training/testing settings.
5. Use the provided scripts to train or test the model.

Visit