Table of Contents

• VideoHallucer
  • Introduction
  • Statistics
  • Data
• VideoHallucerKit
  • Installation
  • Usage
• Leaderboard

Recent advancements in Multimodal Large Language Models (MLLMs) have extended their capabilities to video understanding. Yet, these models are often plagued by "hallucinations", where irrelevant or nonsensical content is generated, deviating from the actual video context. This work introduces VideoHallucer, the first comprehensive benchmark for hallucination detection in large video-language models (LVLMs). VideoHallucer categorizes hallucinations into two main types: intrinsic and extrinsic, offering further subcategories for detailed analysis, including object-relation, temporal, semantic detail, extrinsic factual, and extrinsic non-factual hallucinations. We adopt an adversarial binary VideoQA method for comprehensive evaluation, where pairs of basic and hallucinated questions are crafted strategically. By evaluating eleven LVLMs on VideoHallucer, we reveal that (i) the majority of current models exhibit significant issues with hallucinations; (ii) while scaling datasets and parameters improves models' ability to detect basic visual cues and counterfactuals, it provides limited benefit for detecting extrinsic factual hallucinations; (iii) existing models are more adept at detecting facts than identifying hallucinations. As a byproduct, these analyses further instruct the development of our self-PEP framework, achieving an average of 5.38% improvement in hallucination resistance across all model architectures.

The Extrinsic Factual Hallucination and Extrinsic Non-factual Hallucination share same videos and basic questions

You can download the videohallucer from huggingface, containing both json and videos.

videohallucer_datasets                    
    ├── object_relation
        ├── object_relation.json
        └── videos
    ├── temporal
        ├── temporal.json
        └── videos
    ├── semantic_detail
        ├── semantic_detail.json
        └── videos
    ├── external_factual
        ├── external_factual.json
        └── videos
    └── external_nonfactual
        ├── external_nonfactual.json
        └── videos

We offer a selection of case examples from our dataset for further elucidation:

[
    {
        "basic": {
            "video": "1052_6143391925_916_970.mp4",
            "question": "Is there a baby in the video?",
            "answer": "yes"
        },
        "hallucination": {
            "video": "1052_6143391925_916_970.mp4",
            "question": "Is there a doll in the video?",
            "answer": "no"
        },
        "type": "subject"
    },
...
]

If you want to upload results from your models, feel free to submit a PR following one of these baselines, or send an email to me ([email protected]) to update your results on our page.

Available Baselines

• VideoChatGPT-7B

• Valley2-7B

• Video-LLaMA-2-7B/13B

• VideoChat2-7B

• VideoLLaVA-7B

• LLaMA-VID-7B/13B

• VideoLaVIT-7B

• MiniGPT4-Video-7B

• PLLaVA-7B/13B/34B

• LLaVA-NeXT-Video-DPO-DPO-7B/34B

• ShareGPT4Video-8B

• Gemini-1.5-pro

• GPT4O (Azure)

• LLaVA

• GPT4V (Azure)

For detailed instructions on installation and checkpoints, please consult the INSTALLATION guide.

debug inference pipeline

cd baselines
python ../model_testing_zoo.py --model_name Gemini-1.5-pro # ["VideoChatGPT", "Valley", "Video-LLaMA-2", "VideoChat2", "VideoLLaVA", "LLaMA-VID", "VideoLaVIT", "PLLaVA", "PLLaVA-13B", "PLLaVA-34B", "LLaVA-NeXT-Video", "LLaVA-NeXT-Video-34B", "Gemini-1.5-pro", "GPT4O", "GPT4V", "LLaVA"])

evaluate on VideoHallucer

cd baselines
python ../evaluations/evaluation.py  --model_name Gemini-1.5-pro --eval_obj --eval_obj_rel --eval_temporal --eval_semantic --eval_fact --eval_nonfact

evaluate "yes/no" bias

python ../evaluations/evaluation.py GPT4O Gemini-1.5-pro # ["VideoChatGPT", "Valley", "Video-LLaMA-2", "VideoChat2", "VideoLLaVA", "LLaMA-VID", "VideoLaVIT", "PLLaVA", "PLLaVA-13B", "PLLaVA-34B", "LLaVA-NeXT-Video", "LLaVA-NeXT-Video-34B", "Gemini-1.5-pro", "GPT4O", "GPT4V", "LLaVA"]

more detailed results see baselines/results

• We thank vllm-safety-benchmark for inspiring the framework of VideoHallucerKit.
• We thank Center for AI Safety for supporting our computing needs.

If you find our work helpful, please consider citing it.

@article{videohallucer,
    title={VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models},
    author={Wang, Yuxuan and Wang, Yueqian and Zhao, Dongyan and Xie, Cihang and Zheng, Zilong},
    journal={arxiv},
    year={2024}
}