VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding

2024-12-04

arXiv.org

AI Summary

This work addresses temporal hallucination—a pervasive issue in multimodal large language models (MLLMs) for video understanding—by formally defining and quantifying three hallucination types: action, temporal logic, and scene transition. To systematically evaluate this problem, we introduce VidHalluc, the first large-scale benchmark for video temporal hallucination assessment, comprising 5,002 annotated video pairs; it exposes critical reliability deficiencies in state-of-the-art MLLMs. We propose DINO-HEAL, a training-free mitigation method that leverages DINOv2 to extract robust visual features and applies spatial saliency-guided feature reweighting to suppress hallucinatory outputs. Experiments on VidHalluc demonstrate that DINO-HEAL reduces average hallucination rates by 3.02%. Both the VidHalluc benchmark and the DINO-HEAL implementation are publicly released to foster reproducible research.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Multimodal large language models (MLLMs) have recently shown significant advancements in video understanding, excelling in content reasoning and instruction-following tasks. However, hallucination, where models generate inaccurate or misleading content, remains underexplored in the video domain. Building on the observation that MLLM visual encoders often fail to distinguish visually different yet semantically similar video pairs, we introduce VidHalluc, the largest benchmark designed to examine hallucinations in MLLMs for video understanding. It consists of 5,002 videos, paired to highlight cases prone to hallucinations. VidHalluc assesses hallucinations across three critical dimensions: (1) action, (2) temporal sequence, and (3) scene transition. Comprehensive testing shows that most MLLMs are vulnerable to hallucinations across these dimensions. Furthermore, we propose DINO-HEAL, a training-free method that reduces hallucinations by incorporating spatial saliency from DINOv2 to reweight visual features during inference. Our results show that DINO-HEAL consistently improves performance on VidHalluc, achieving an average improvement of 3.02% in mitigating hallucinations across all tasks. Both the VidHalluc benchmark and DINO-HEAL code are available at https://people-robots.github.io/vidhalluc.

Problem

• Evaluating hallucinations in MLLMs for video understanding
• Assessing model inaccuracies in action, sequence, and scene transitions
• Proposing a method to reduce hallucinations in video MLLMs

Innovation

• Largest benchmark VidHalluc for video hallucinations
• Training-free DINO-HEAL reduces hallucinations effectively
• Uses DINOv2 spatial saliency to reweight features