EventHallusion

EventHallusion is the first benchmark that focuses on the evaluation of event hallucinations in Video LLMs. This repository includes the code and benchmark data of our paper "EventHallusion: Diagnosing Event Hallucinations in Video LLMs", by Jiacheng Zhang, Yang Jiao, Shaoxiang Chen, Jingjing Chen, and Yu-Gang Jiang.

Update

• [2024/09/24] The GitHub repo of EventHalluion is created. The code and data for evaluation are released.
• [2024/09/27] The paper is released on Arxiv.

Introduction

Recently, Multimodal Large Language Models (MLLMs) have made significant progress in the video comprehension field. Despite remarkable content reasoning and instruction following capabilities they demonstrated, the hallucination problem of these VideoLLMs is less explored compared with its counterpart in the image domain. To mitigate this gap, we first propose EventHallusion, a novel benchmark that focuses on assessing the VideoLMMs' hallucination phenomenon on video event comprehension. Based on the observation that existing VideoLLMs are entangled with the priors stemming from their foundation models, our EventHallusion is curated by meticulously collecting videos and annotating questions to intentionally mislead the VideoLLMs into interpreting events based on these priors rather than accurately understanding the video content. On the other hand, we also propose a simple yet effective method, called Temporal Contrastive Decoding (TCD), to tackle the hallucination problems of VideoLLMs. The proposed TCD suppresses the model's preference toward their priors by comparing the original video with a constructed counterpart, whose temporal cues are disrupted, during the autoregressive decoding stage. Through comprehensive evaluation of eight open-source and two closed-source VideoLLMs on the proposed EventHallusion benchmark, we find that the open-source models suffer significantly from hallucination problems, whereas the closed-source models perform markedly better. By further equipping open-sourced VideoLLMs with the proposed TCD approach, evident performance improvements are achieved across most metrics in the EventHallusion benchmark

Getting Started

Prepare Environment and Data

a. Clone this repository.

git clone https://github.com/Stevetich/EventHallusion.git
cd EventHallusion/

b. Create the environment by conda and install necessary packages.

conda create -n eventhallusion python=3.10 -y
conda activate eventhallusion
pip install -r requirements.txt

c. Prepare Data.

We have released our videos, which can be downloaded from google drive. The questions for evaluation can be found in questions/.

Run inference

1. GPT-4o We provide scripts for inference with gpt-4o. First, input your own openai api key in run_gpt4o.py file, and run with:

python run_gpt4o.py --video_path <downloaded_video_path>

2. Video LLaVA We also provide a simple example of evaluating Video-LLaVA. We recommend to follow the official installation in Video-LLaVA or run the following commands for simplicity:

cd inference_template/
pip install -e .
pip install decord opencv-python git+https://github.com/facebookresearch/pytorchvideo.git@28fe037d212663c6a24f373b94cc5d478c8c1a1d

After installation, download the checkpoint of Video-LLaVA in Video-LLaVA and run inference with the following commands:

cd videollava/eval/self_bench/
python run_inference.py --model_path <your_path_to_video_llava_ckpt> --video_path <downloaded_video_path>

Besides, we provide examples of prediction files, which can be found in predictions/

Run Evaluation

After obtaining the prediction results file, run the following commands for metrics:

python eval.py --input_file <your_path_to_prediction_json>

Citation

If you find our paper useful, please cite:

@article{eventhallusion,
    title={EventHallusion: Diagnosing Event Hallucinations in Video LLMs},
    author={Jiacheng Zhang, Yang Jiao, Shaoxiang Chen, Jingjing Chen, Yu-Gang Jiang},
    journal={arXiv preprint arXiv: 2409.16597},
    year={2024}
}

Acknowledgement

We sincerely thank authors of Video-LLaVA, TempCompass, VideoHallucer for opening their methods.