Datasets:

SEU-VIPGroup
/

FG-BMK

@@ -1,201 +1,76 @@
-<div align="center">
-<h1>
-  <img src="static/images/benchmark.svg" width="40" align="absmiddle" alt=""/>&nbsp; FG-BMK
-</h1>
-<br/>
-<a href="https://arxiv.org/abs/2606.19053">
-  <img src="https://img.shields.io/badge/arXiv%20(EXTEND)-2606.19053-b31b1b.svg?style=for-the-badge&logo=arxiv&logoColor=white" alt="arXiv EXTEND version"/>
-</a>
-<a href="https://arxiv.org/abs/2504.14988">
-  <img src="https://img.shields.io/badge/arXiv%20(ICLR)-2504.14988-b31b1b.svg?style=for-the-badge&logo=arxiv&logoColor=white" alt="arXiv ICLR version"/>
-</a>
-<a href="https://fg-bmk.github.io/">
-  <img src="https://img.shields.io/badge/Project-Page-0f6e63.svg?style=for-the-badge&logo=githubpages&logoColor=white" alt="Project Page"/>
-</a>
-<a href="https://github.com/SEU-VIPGroup/FG-BMK">
-  <img src="https://img.shields.io/badge/Code-GitHub-181717.svg?style=for-the-badge&logo=github&logoColor=white" alt="Code"/>
-</a>
-</div>
 ---
-> 📄 This repository contains the **data** and **evaluation code** for **FG-BMK**.
-> For an interactive leaderboard, dataset overview, and key findings, visit the **[project page](https://fg-bmk.github.io/)**.
-## 🔔 News
-- 🔥 (***new***) An extended version of our paper is now on [arXiv](https://arxiv.org/abs/2606.19053).
-- ✨ Major benchmark update — we added new dataset and expanded the suite with additional experimental validation and analysis.
-- 🎉 Our paper has been **accepted to ICLR 2026**! The ICLR version is available on [arXiv](https://arxiv.org/abs/2504.14988).
-- 🚀 We released our **FG-BMK** benchmark!
-## 📖 Introduction
-<p align="center">
-  <img src="FG-BMK.png" width="92%" alt="FG-BMK overview"/>
-</p>
-**FG-BMK** is a comprehensive fine-grained evaluation benchmark — **1.01 million questions** over **0.28 million images** — providing a robust test bed for evaluating LVLMs. It incorporates two complementary evaluation paradigms:
-- **🧑 Human-oriented evaluation** — employs dialogue-like interactions to assess a model's ability to understand and respond to fine-grained visual queries in a conversational context.
-- **🤖 Machine-oriented evaluation** — focuses on two core fine-grained vision tasks, **image retrieval** and **recognition**, to directly measure the feature-representation capabilities of LVLMs.
-Compared with existing efforts — which primarily focus on fine-grained classification or use a limited number of questions — FG-BMK enables a comprehensive assessment of LVLMs' fine-grained feature representation and semantic recognition abilities. Our evaluations of **eight open-source LVLMs/VLMs** uncover key findings on the influence of training paradigms, modality alignment, perturbation susceptibility, and fine-grained category reasoning on task performance.
 ---
-# 🚀 Evaluation Guidelines
-## 1. Dataset Preparation
-Before running inference, download the corresponding dataset images. Links to the source dataset projects are provided below — you can download the datasets to any location:
-| Dataset | Dataset | Dataset |
-|:--|:--|:--|
-| [FGVC-Aircraft (aircraft)](https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/) | [CUB-200-2011 (cub)](https://www.vision.caltech.edu/datasets/cub_200_2011/) | [DeepFashion (deepfashion)](http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion.html) |
-| [Oxford 102 Flower (flowers102)](https://www.robots.ox.ac.uk/~vgg/data/flowers/102/) | [Food-101 (food101)](https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/) | [iNat2021 (iNat2021)](https://github.com/visipedia/inat_comp/tree/master/2021) |
-| [Products-10K (products10k)](https://products-10k.github.io/challenge.html#downloads) | [SkinCon (skincon)](https://skincon-dataset.github.io/) | [Stanford Car (stanfordcar)](https://www.sighthound.com/products/alpr) |
-| [Stanford Dog (stanforddog)](http://vision.stanford.edu/aditya86/ImageNetDogs/) | [VegFru (vegfru)](https://github.com/ustc-vim/vegfru) | [Wine (wine)](https://tianchi.aliyun.com/dataset/110147) |
-| [MTARSI (mtarsi)](https://huggingface.co/datasets/amistele/MTARSI-fixed) | | |
-- **Human-oriented evaluations** — questions are pre-generated for each image, e.g. `benchmark/human-oriented/attribute_recognition/cub_attribute_questions.jsonl`.
-- **Machine-oriented evaluations** — dataset categories and the corresponding `train.csv` / `test.csv` splits live in directories such as `benchmark/machine-oriented/aircraft`.
-## 2. Environment & Checkpoint Preparation
-Our evaluation covers **eight LVLMs/VLMs**. The table below lists the project and checkpoint for each model, to help you configure the environment and download the relevant weights:
-| Model Project | Checkpoint |
-|:--|:--|
-| [InternVL](https://github.com/OpenGVLab/InternVL) | [InternVL-Chat-V1.1](https://huggingface.co/OpenGVLab/InternVL-Chat-V1-1) |
-| [LLaVA-1.5](https://github.com/haotian-liu/LLaVA) | [LLaVA-1.5-7B](https://huggingface.co/liuhaotian/llava-v1.5-7b) |
-| [Qwen-VL](https://github.com/QwenLM/Qwen-VL) | [Qwen-VL-Chat](https://huggingface.co/Qwen/Qwen-VL-Chat) |
-| [BLIP-2](https://github.com/salesforce/LAVIS/tree/main/projects/blip2) | [BLIP-2-FLAN-T5-XL](https://huggingface.co/Salesforce/blip2-flan-t5-xl) |
-| [EVA-CLIP](https://github.com/baaivision/EVA/tree/master/EVA-CLIP) | [EVA02_CLIP_L_psz14_s4B](https://huggingface.co/QuanSun/EVA-CLIP/blob/main/EVA02_CLIP_L_psz14_s4B.pt) |
-| [BEiT3](https://github.com/microsoft/unilm/blob/master/beit3/README.md) | [BEiT3-large-itc](https://github.com/addf400/files/releases/download/beit3/beit3_large_itc_patch16_224.pth) |
-| [CoCa](https://github.com/mlfoundations/open_clip#fine-tuning-coca) | [CoCa-L](https://github.com/mlfoundations/open_clip#fine-tuning-coca) |
-| [DINOv2](https://github.com/facebookresearch/dinov2) | [DINOv2-L](https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_pretrain.pth) |
-## 3. Inference
-### 🧑 Human-oriented Evaluation
-To use your own model and produce final answers, first modify the model-loading code in `human_evaluation_demo.py` to adapt it to your model. Here is an example of loading the **InternVL** model:
-```python
-# Load InternVL model, tokenizer, and image processor
-from transformers import AutoModel, AutoTokenizer, CLIPImageProcessor
-model = AutoModel.from_pretrained(
-    args.model_path,
-    torch_dtype=torch.bfloat16,
-    low_cpu_mem_usage=True,
-    use_flash_attn=True,
-    trust_remote_code=True).eval().cuda()
-tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True, use_fast=False)
-image_processor = CLIPImageProcessor.from_pretrained(args.model_path)
-```
-The model then answers the questions through its inference code, e.g.:
-```python
-# Load images
-image = Image.open(image_path).resize((448, 448))
-pixel_values = image_processor(images=image, return_tensors='pt').pixel_values.to(torch.bfloat16).cuda()
-# Generate response
-generation_config = dict(max_new_tokens=1024, do_sample=True)
-response = model.chat(tokenizer, pixel_values, prompt_text, generation_config)
-```
-After modifying the loading code, configure **`model-path`** (checkpoint), **`question-file`**, **`image-folder`** (path to the dataset), and **`answers-file`** (output path) in `run_human_demo.sh`, then run:
-```bash
-bash run_human_demo.sh
-# The code splits the question file by the number of GPUs and runs inference concurrently.
-```
-The outputs are merged into a single file in the following format:
-```json
-{"question_id": 1, "image": "images/001.Black_footed_Albatross/Black_Footed_Albatross_0078_796126.jpg", "prompt": "Is the genus of the object geococcyx? Answer with yes or no.", "text": "No", "class": "no", "category": "generic"}
-{"question_id": 2, "image": "images/001.Black_footed_Albatross/Black_Footed_Albatross_0003_796136.jpg", "prompt": "Is the genus of the object raven? Answer with yes or no.", "text": "No", "class": "no", "category": "generic"}
-```
-Finally, compute accuracy with `answer_acc.py`:
-```bash
-python answer_acc.py
-```
-> See the [example output](https://github.com/MrPetrichor/FG-BMK/blob/main/demo/human_evaluation/example_output.jsonl) for a detailed prediction-file format. We also provide reference inference code for **Qwen-VL** and **BLIP-2**.
-### 🤖 Machine-oriented Evaluation
-To evaluate an LVLM's feature-representation ability, first modify the feature-extraction code in `models.py`. Here is an example of **CoCa** feature extraction:
-```python
-def coca(model_name, pretrained, cache_dir):
-    from open_clip import create_model_and_transforms
-    def _hook(self, _, input, output):
-        self.feat.append(output.transpose(0, 1))
-    def get_intermediate_layers(self, x, n=1, return_class_token=True):
-        self.feat = []
-        self(x)
-        class_tokens = [out[:, 0] for out in self.feat]
-        outputs = [out[:, 1:] for out in self.feat]
-        return tuple(zip(outputs, class_tokens))
-    model, _, preprocess = create_model_and_transforms(model_name, pretrained, cache_dir=cache_dir)
-    model = model.visual
-    model.eval()
-    model.cuda()
-    model.__class__._hook = _hook
-    model.__class__.get_intermediate_layers = get_intermediate_layers
-    model.transformer.resblocks[-2].register_forward_hook(model._hook)
-    model.transformer.resblocks[-1].register_forward_hook(model._hook)
-    return model
-```
-In this module we use **`_hook`** and the defined **`get_intermediate_layers`** method to extract visual features from the last two layers of the vision encoder. We then concatenate the **cls token** and **image tokens** in the predefined order and return an instance of the **CoCa model**. Examples of visual feature extraction using **EVA-CLIP** and **Qwen-VL** are already provided in `model.py`.
-Once modified, import the model in `eval_linear.py` or `eval_retrieval.py`:
-```python
-from dinov2.utils.config import setup
-from models import coca
-torch.backends.cudnn.benchmark = True
-model = coca('coca_ViT-L-14', 'laion2b_s13b_b90k', '.cache')
-config = setup(args)
-autocast_dtype = torch.float16
-```
-Then run the demo:
-```bash
-python eval_linear.py
-# or
-python eval_retrieval.py
-```
-The output log will look like:
-```
-I20240504 13:34:23 16157 dinov2 helpers.py:103] Training  [    0/10000]  eta: 8:13:21  loss: 143.1147 (143.1147)  lr: 0.0005 (0.0005)  time: 2.960182  data: 2.449736  max mem: 2711
-...
-```
-After the training process completes, the code automatically runs inference on the test set of the fine-grained dataset and reports the results.
----
-## 📌 Citation
-If you find FG-BMK useful for your research, please consider citing:
 ```bibtex
 @article{yu2026fgbmk,
@@ -206,6 +81,6 @@ If you find FG-BMK useful for your research, please consider citing:
 }
 ```
-<div align="center">
-<sub>For questions or issues, please open a <a href="https://github.com/SEU-VIPGroup/FG-BMK/issues">GitHub issue</a>.</sub>
-</div>

 ---
+pretty_name: FG-BMK
+license: other
+language:
+- en
+task_categories:
+- image-classification
+- image-retrieval
+- visual-question-answering
+tags:
+- fine-grained-recognition
+- large-vision-language-models
+- benchmark
+- image-retrieval
+- visual-question-answering
+size_categories:
+- 100K<n<1M
 ---
+# 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](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-Aircraft) | `aircraft.zip` |
+| CUB-200-2011 | [SEU-VIPGroup/FG-BMK-image-CUB](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-CUB) | `CUB.zip` |
+| DeepFashion | [SEU-VIPGroup/FG-BMK-image-DeepFashion](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-DeepFashion) | `deepfashion.tar` |
+| Oxford 102 Flowers | [SEU-VIPGroup/FG-BMK-image-Flowers102](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-Flowers102) | `flowers102.tar` |
+| Food-101 | [SEU-VIPGroup/FG-BMK-image-Food101](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-Food101) | `food101.tar` |
+| iNat2021 validation images | [SEU-VIPGroup/FG-BMK-image-Val](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-Val) | `val.tar.gz` |
+| SkinCon | [SEU-VIPGroup/FG-BMK-image-SkinCon](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-SkinCon) | `skincon.tar` |
+| Stanford Dogs | [SEU-VIPGroup/FG-BMK-image-Dog](https://huggingface.co/datasets/SEU-VIPGroup/FG-BMK-image-Dog) | `dog.zip` |
+| VegFru | [SEU-VIPGroup/FG-BMK-image-VegFru](https://huggingface.co/datasets/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
 ```bibtex
 @article{yu2026fgbmk,
 }
 ```
+## 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.