🔬 FD²: A Dedicated Framework for Fine-Grained Dataset Distillation (ECCV 2026)

FD² is a dedicated dataset distillation framework for fine-grained recognition. It localizes discriminative regions, constructs fine-grained representations, and preserves diverse class-specific characteristics in distilled datasets.

📢 News

• The code is available in this repository.
• June 2026: Our paper has been accepted to ECCV 2026!
• March 2026: Preprint was released.

🎯 Key Contributions

• A Dedicated Framework for Fine-Grained Dataset Distillation Introduces a framework specifically designed for fine-grained recognition tasks, where subtle inter-class differences and large intra-class variations must be carefully preserved.

• Counterfactual Attention Learning Identifies localized discriminative regions during network pretraining and aggregates their representations to construct and update fine-grained class prototypes.

• Fine-Grained Characteristic Constraint Aligns each distilled sample with its corresponding class prototype while repelling it from other class prototypes, improving inter-class separability and preserving class-specific characteristics.

• Similarity Constraint for Intra-Class Diversity Encourages different distilled samples from the same class to attend to complementary discriminative regions, preventing sample homogenization and preserving diverse fine-grained visual cues.

• Seamless Integration and Strong Transferability Integrates readily into existing decoupled dataset distillation pipelines and consistently improves performance across fine-grained and general image classification datasets.

⚙️ Installation

This project is built on CV-DD. Please refer to the CV-DD repository for environment setup and installation instructions.

🧰 Data Preparation Utilities

• RDED_patch.py generates patch-based initialization samples from the original training set using a pretrained classifier.
• get_small_ipc_from_big_ipc.py transfers distilled images from a larger IPC setting to a smaller one. After generating the IPC=5 distilled set, set target_ipc to 3 or 1 and update the source and target paths to construct the corresponding smaller distilled set.

📊 Main Results

FD² consistently improves SRe²L++ and FADRM+ on CUB-200-2011, FGVC Aircraft, and Stanford Cars across IPC settings of 1, 3, and 5. The improvements are especially pronounced on FGVC Aircraft and Stanford Cars, demonstrating the value of preserving localized and class-specific fine-grained cues.

🖼️ Visualization of Distilled Samples

Compared with the corresponding baselines, integrating FD² produces distilled samples with clearer local structures, richer texture details, and more discriminative class-specific cues.

📚 Citation

If you find this project useful for your research, please use the following BibTeX entry.

@inproceedings{ma2026fd2,
  title={{FD}$^2$: A Dedicated Framework for Fine-Grained Dataset Distillation},
  author={Ma, Hongxu and Li, Guang and Wang, Shijie and Zhou, Dongzhan and Sun, Baoli and Ogawa, Takahiro and Haseyama, Miki and Wang, Zhihui},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2026}
}