Learning Separable Fine-Grained Representation via Dendrogram Construction from Coarse Labels for Fine-grained Visual Recognition

This repo contains PyTorch implementation of the BucSFR. BucSFR is a method that learns fine-grained representations form coarsely labeled datasets without any supervision at the fine-grained level. For more details please check our paper BucSFR.

---

Installation

Setup the conda environment and install the required packages by running the following commands:

conda env create -n BuCSFR -f requirements.txt
conda activate BuCSFR

---

Dataset Preparation

• For ImageNet32, we followed the repo MaskCon;
• For CIFAR100, CIFAR10 and Stanford Cars, it will download the datasets automatically in your input path;
• For iNaturalist2019, we followed this repo Making Better Mistakes;
• For FGVC datasets, please download them from the official website:
  • FGVC-Aircraft
  • CUB200-2011
  • Oxford Flowers102
  • NABirds

---

Training

To train BuCSFR , please run the following command in the train.sh. For example, to train BucSFR on the CIFAR100 dataset, run:

CUDA_VISIBLE_DEVICES=0,3 python main.py --dist-url tcp://localhost:10009 --multiprocessing-distributed --world-size 1 --rank 0 
--dataset cifar100 --arch resnet50 --img_size 224 --lr 0.03 --batch-size 256 --moco-k 65536 --moco-t 0.2 
--data 'your data path' 
--exp_dir ./experiment/cifar100 
--warmup_epoch 10 --epochs 100 --workers 8 --mlp --aug-plus --cos 
| tee -a zz_cifar100.log 

---

Citation

If you find our code useful, please consider citing:

@InProceedings{Shi_2025_ICCV,
    author    = {Shi, Guanghui and Liang, Xuefeng and Li, Wenjie and Lin, Xiaoyu},
    title     = {Learning Separable Fine-Grained Representation via Dendrogram Construction from Coarse Labels for Fine-grained Visual Recognition},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {870-879}
}.