FG-BMK

FG-BMK is a fine-grained evaluation benchmark for large vision-language models and vision-language models. It contains benchmark metadata, generated questions, train/test splits, class lists, and evaluation code for studying fine-grained visual understanding from both semantic and feature-representation perspectives.

The benchmark covers 1.01 million questions over 0.28 million images. It evaluates two complementary settings:

• Human-oriented evaluation: dialogue-style fine-grained visual questions, including hierarchical category recognition, attribute recognition, and knowledge-bias estimation.
• Machine-oriented evaluation: fine-grained recognition and image retrieval using predefined train/test splits and class files.

This repository contains the benchmark files and evaluation code. Image archives are hosted separately as companion Hugging Face dataset repositories.

Companion Image Repositories

Dataset subset	Repository	Archive
FGVC-Aircraft	SEU-VIPGroup/FG-BMK-image-Aircraft	aircraft.zip
CUB-200-2011	SEU-VIPGroup/FG-BMK-image-CUB	CUB.zip
DeepFashion	SEU-VIPGroup/FG-BMK-image-DeepFashion	deepfashion.tar
Oxford 102 Flowers	SEU-VIPGroup/FG-BMK-image-Flowers102	flowers102.tar
Food-101	SEU-VIPGroup/FG-BMK-image-Food101	food101.tar
iNat2021 validation images	SEU-VIPGroup/FG-BMK-image-Val	val.tar.gz
SkinCon	SEU-VIPGroup/FG-BMK-image-SkinCon	skincon.tar
Stanford Dogs	SEU-VIPGroup/FG-BMK-image-Dog	dog.zip
VegFru	SEU-VIPGroup/FG-BMK-image-VegFru	vegfru.zip

Other subsets referenced by the benchmark should be obtained from their original dataset sources when they are not mirrored in the companion repositories.

Repository Contents

• benchmark/human-oriented/: question files for human-oriented LVLM evaluation.
• benchmark/machine-oriented/: class lists and train/test splits for recognition and retrieval evaluation.
• demo/human_evaluation/: example inference and answer-scoring code.
• demo/machine_evaluation/: example feature extraction, linear evaluation, and retrieval evaluation code.
• static/ and project assets: figures and supporting files for the benchmark release.

Basic Use

1. Download this repository.
2. Download the needed image archive from the companion repository above.
3. Extract the archive locally.
4. Point the evaluation scripts to the extracted image folder and to the corresponding FG-BMK question or split file.

For human-oriented evaluation, use files under benchmark/human-oriented/ and run the demo code in demo/human_evaluation/.

For machine-oriented evaluation, use files under benchmark/machine-oriented/<dataset>/ and run the demo code in demo/machine_evaluation/.

Paper and Code

• Project page: https://fg-bmk.github.io/
• Code repository: https://github.com/SEU-VIPGroup/FG-BMK
• arXiv extended version: https://arxiv.org/abs/2606.19053
• arXiv ICLR version: https://arxiv.org/abs/2504.14988

Citation

@article{yu2026fgbmk,
  title   = {Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: From Evaluation to Diagnosis},
  author  = {Yu, Hong-Tao and Xie, Chen-Wei and Peng, Yuxin and Belongie, Serge and Wei, Xiu-Shen},
  journal = {arXiv preprint arXiv:2606.19053},
  year    = {2026}
}

@inproceedings{yu2026fgbmk_iclr,
  title     = {Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation},
  author    = {Yu, Hong-Tao and Peng, Yuxin and Belongie, Serge and Wei, Xiu-Shen},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

License and Data Terms

The benchmark metadata and evaluation code follow the terms provided by the project authors. Image archives are derived from their corresponding source datasets; users must also follow the license and usage terms of each original dataset.