R-Bench

This repo is based on the previous repo of the paper Evaluating and Analyzing Relationship Hallucinations in Large Vision-Language Models (ICML2024) and our own code.

@InProceedings{pmlr-v235-wu24l,
  title = 	 {Evaluating and Analyzing Relationship Hallucinations in Large Vision-Language Models},
  author =       {Wu, Mingrui and Ji, Jiayi and Huang, Oucheng and Li, Jiale and Wu, Yuhang and Sun, Xiaoshuai and Ji, Rongrong},
  booktitle = 	 {Proceedings of the 41st International Conference on Machine Learning},
  pages = 	 {53553--53570},
  year = 	 {2024},
  editor = 	 {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix},
  volume = 	 {235},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {21--27 Jul},
  publisher =    {PMLR},
  pdf = 	 {https://raw.githubusercontent.com/mlresearch/v235/main/assets/wu24l/wu24l.pdf},
  url = 	 {https://proceedings.mlr.press/v235/wu24l.html},
  }

Data

Download R-Bench. The main annotation files include:

- image-level_filterd.json
- instance-level_filterd.json
- nocaps_pope_obj_random_image.json
- nocaps_pope_obj_popular_image.json
- nocaps_pope_obj_adversarial_image.json
- web_data
- instance_mask

These files contain annotations for image-level, instance-level(box and mask share same questions), pope-object, and web-data questions. For image-level and instance-level questions, we randomly sampled five subsets into the [type]_ids_[subset].json files.

Download the images from image (source from Open Image validation set (v4)).

wget https://drive.google.com/file/d/10JXRdzTRMylWQA160qdoYITGO10g6N8k/view?usp=sharing

unzip image.zip

Eval

Step1: To run LVLM on R-Bench using the official inference script of the LVLMs.

For Image-level, pseudocode is as follows,

for line in questions:
    question_id = line['question_id']
    question = line['text']
    image = open(line['image'])
    text = model(question, image)
    answer_file.write(json.dumps("question_id": question_id, "text":text))

For Instance-level, pseudocode is as follows, set instance_level_box=True or instance_level_mask=True to get result for Instance-Level result with Box or Mask.

import instance_qs_construct, draw_box, draw_mask

for line in questions:
    question_id = line['question_id']
    question = instance_qs_construct(line, type='mask' if instance_level_mask else 'box')
    if instance_level_box:
      image = draw_box(line)
    elif instance_level_mask:
      image = draw_mask(line)
    text = model(question, image)
    answer_file.write(json.dumps("question_id": question_id, "text":text))

and format the result.json file as follows:

{"question_id": 0, "text":[model output]}
{"question_id": 1, "text":[model output]}
...

Tips: We provide instance-level question tools in utils.py. Please use the draw_mask and draw_box functions to draw the mask or box on input images, respectively. Additionally, use the instance_qs_construct function to reformat the instance questions.

Step 1.1: Evaluate with open-source LLaVA. We provide LLaVA demos for reference. After setting up the LLaVA environment, place the demos/eval-rbench.py file into the llava/eval/ directory and execute llava-r-bench.sh.

# 1. Clone the repository and navigate to LLaVA folder
git clone https://github.com/haotian-liu/LLaVA.git
cd LLaVA

# 2. Install Package
conda create -n llava python=3.10 -y
conda activate llava
pip install --upgrade pip  # enable PEP 660 support
pip install -e .

# 3. Copy the eval-rbench.py and lllava-r-bench.sh to LLaVA folder
cp R-bench/demos/eval-rbench.py LLaVA/llava/eval/
cp R-bench/demos/llava-r-bench.sh LLaVA/

# 4. Run the script
chmod +x llava-r-bench.sh
./llava-r-bench.sh

Step 1.2: Evaluate with API. We provide an example of using ZhipuAI's API for evaluation. Here's how to get started:

1. Install the required package:

pip install zhipuai

2. Set up your API key:

import zhipuai
zhipuai.api_key = "YOUR_API_KEY"

3. Run the evaluation script:

python test_glm_api.py  # default: image-level evaluation

You can specify different evaluation types using the --type argument:

# For image-level evaluation
python test_glm_api.py --type image-level

# For instance-level evaluation with masks
python test_glm_api.py --type instance-level-mask

# For instance-level evaluation with boxes
python test_glm_api.py --type instance-level-box

The script will process the images and questions according to the specified type and save the results in the format:

{"question_id": 0, "qtype": "positive", "answer": "[model output]"}
{"question_id": 1, "qtype": "negative", "answer": "[model output]"}
...

Note: Make sure to properly configure your API key and handle rate limits according to your API usage agreement.

Step2: Eval with,

sh eval.sh

Tips: You can just replace --result-file with the file you generated, and the eval.py script will automatically calculate the average results for different subsets. The results obtained through our script are for the Image(All), and the results for Image(subset) in the paper are for reference only.

Step3: Fine-grained reasoning evaluation.

To evaluate the quality of reasoning explanations generated by the model, we compare them with human-annotated "ground truth" explanations using cosine similarity between their sentence embeddings.

1. Install Dependencies:

pip install sentence-transformers scikit-learn numpy

2. Prepare Data: Ensure your result file (e.g., result.json) contains entries in the following format:

{
    "question id": 6187,
    "answer": "Model generated explanation",
    "labeled reason": "Human annotated ground truth"
}

3. Run Evaluation:

python cal_cos_similarity.py

The evaluation script will output:

• Cosine similarity score for each question
• Average similarity score across all questions

Note: By default, we use all-MiniLM-L6-v2 as the sentence encoder, which is a lightweight yet effective model. You can modify the model name in the calculate_cosine_similarity function if you prefer to use other pre-trained models.

分工

黄叙川：

1. 阅读论文、准备presentation PPT、presentation
2. 复现论文中测试LLaVA的结果
3. 基于论文已有benchmark调用glm-4v-flash的API进行测试、
4. 撰写论文
5. 考虑R-Bench中欠缺考虑的情况，构思test cases进行补充

吕嘉楠：

1. 阅读论文，准备pre和ppt
2. 在glm-4v-plus-0111和kimi上测试
3. 撰写论文
4. 构建subset寻找原始R-bench存在的问题

Acknowledge

The evaluation code is based on POPE.