Logo VideoHallucer

Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models

Yuxuan Wang1,2, Yueqian Wang2,3, Dongyan Zhao2,3, Cihang Xie3, Zilong Zheng1,2,,
1Beijing Institute for General Artificial Intelligence,
1State Key Laboratory of General Artificial Intelligence,
2Peking University, 3UC, Santa Cruz

arXiv   Code   Data  

Introduction

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.

Leaderboard

Accuracy scores on the VideoHallucer.

# Model LLM Frames Date Yes/No Bias Accuracy
Pct. Diff (∼ 0) FP Ratio (∼ 0.5) Basic Hallucinated Overall
1 Human - - 2024-05-25 0.02 0.42 90 88.8 85
2 GPT-4o - 16 2024-06-18 -0.02 0.43 75.1 74.2 53.3
3 PLLaVA-34B Yi-34B 16 2024-05-25 0.18 0.78 90.8 50.8 45
4 PLLaVA-13B Vicuna-13B-1.5 16 2024-05-25 0.17 0.72 87.5 48.6 41.2
5 PLLaVA Vicuna-7B-1.5 16 2024-05-25 0.06 0.53 75.1 55.5 38.1
6 Gemini 1.5 Pro - 1 fps (-128) 2024-05-25 0.15 0.62 83.6 42.3 37.8
7 LLaMA-VID-13B Vicuna-13B-v1.5 1 fps 2024-05-25 0.21 0.72 85.2 36.9 29.2
8 LLaVA-NeXT-Video-DPO-34B Yi-34B 4 2024-05-25 0.07 0.55 73.6 51.6 32.3
9 LLaVA-NeXT-Video-DPO Vicuna-7B-1.5 4 2024-05-25 -0.04 0.40 62.5 60.9 32.0
10 MiniGPT4-Video Mistral-7B (-45) 2024-05-25 0.18 0.62 79.4 28.6 22.3
11 LLaMA-VID Vicuna-7B-v1.5 1 fps 2024-05-25 0.29 0.83 89.9 26.6 21
12 VideoLaVIT LLaMA2-7B 16 2024-05-25 0.36 0.91 94.9 21.3 18.9
13 Video-LLaVA Vicuna-7B-v1.5 8 2024-05-25 0.36 0.91 95.1 20.3 17.8
14 ShareGPT4Video LLaMA3-8B 16 2024-06-25 0.31 0.79 88.5 20.0 15.8
15 Video-LLaMA2 LLaMA2-7B 8 2024-05-25 0.36 0.84 90.9 12.7 10
16 VideoChat2 Vicuna-7B-v0 4 2024-05-25 -0.24 0.15 29.7 25.8 7.8
17 VideoChatGPT LLaMA-7B 100 2024-05-25 0.40 0.89 92.8 10.4 6.4
18 Video-LLaMA2-13B LLaMA2-13B 8 2024-05-25 0.36 0.79 88.3 3.8 3.3
19 Valley2 LLaMA2-7B 8 2024-05-25 -0.07 0.29 44.4 11.5 2.8

Accuracy scores on the different VideoHallucer settings.

Logo VideoHallucer Dataset

Statistics

Statistic Object-Relation Temporal Semantic-Detail Extrinsic-Factual Extrinsic-Nonfactual All
Questions 200 200 200 200 200 1800
Videos 183 165 400 200 200 948
Avg. Question Length 23.7 69.2 27.3 92.3 94.7 61.4
Avg. Video Length (s) 7.0 33.8 13.5 187.0 187.0 85.6

Data Samples

Experiment Results

BibTeX

@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}
}