Fine-Grained Few-Shot Classification with Part Matching

Official code for the FGVC 2025 workshop paper:
📄 Fine-Grained Few-Shot Classification with Part Matching

This repository contains the official implementation of Simple Matching Parts Learner (SMPL), our proposed method for fine-grained few-shot classification using part-level feature matching. We currently provide code to replicate our experiments on the Hotels-8K dataset, with additional datasets to be released in the future.

---

📦 Downloads

🔹 Precomputed Part Features

We provide precomputed part features for the Hotels-8K dataset using either DINO or DINOv2 as the Part Encoder:

➡️ Dino

➡️ DinoV2

Extract to:

After downloading, extract the file to the following path:

data/
│   └── {dataset}/
│       └── features/
│           └── {dataset}_{part_encoder}.npy

• {dataset}: e.g., hotels
• {part_encoder}: either dino or dinov2

This .npy file is a serialized Python dictionary containing part features for the training, validation, and test sets. Each entry maps an image path to a nested dictionary with two keys:

• part_features → a NumPy array of shape (N + 1, D), where N is the number of distinct parts in the image and D is the feature dimensionality. The last row is the global image feature.
• part_indices → a NumPy array of shape (N + 1,), where each value corresponds to the part ID in the segmentation mask. The final value (for the global feature) is always 1000.

🖼️ Original Images & Segmentations

You can download the original images and segmentation maps (generated using GroundedSAM) here:

➡️ Download Images and Segmentations

---

🚀 Training The Part Matcher

To train the Part Matcher model, run:

python train.py –data_dir data –dataset {dataset} –part_encoder {dino|dinov2}

Training episodes are generated stochastically, sampling between 5–20 ways and 1–5 shots per episode, as described in the paper. Class splits used for episode generation are located at:

data/{dataset}/splits/

By default, the best checkpoint will be saved to:

checkpoints/{dataset}/{part_encoder}/best.pt

📊 Evaluation

To evaluate on 20-way classificatio tasks, run:

python evaluation.py 
–dataset {dataset} 
–part_encoder {dino|dinov2} 
–checkpoint {path_to_model} 
–n_shots {1|5}

Evaluation episodes are pre-generated and stored at:

data/{dataset}/test_episodes/20way_{n_shots}shots.json

📖 Citation

If you use our work, please cite:

@InProceedings{Black_2025_CVPR,
    author    = {Black, Samuel and Souvenir, Richard},
    title     = {Fine-grained Few-Shot Classification with Part Matching},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops},
    month     = {June},
    year      = {2025},
    pages     = {2057-2067}
}