Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comprehensive Analysis for Visual Object Hallucination in Large Vision-Language Models

Liqiang Jing, Guiming Hardy Chen, Ehsan Aghazadeh, Xin Eric Wang, Xinya Du

TL;DR

This work performs a component-level analysis of visual object hallucination in LVLMs, decomposing sources of error into the LLM, vision backbone, and projector. It demonstrates that LLMs can faithfully describe content when provided accurate textual captions, while hallucinations primarily emanate from the vision encoder and projector. The authors propose targeted mitigations: fine-grained CLIP tuning and perception-based visual instruction tuning to strengthen the vision backbone, and contrastive alignment losses to improve projector alignment, alongside two fine-grained benchmarks (QA-VisualGenome and QA-FB15K) to evaluate perception and cognition hallucinations. The results show meaningful reductions in perception-based hallucinations and gains in cognition-based tasks, highlighting practical improvements for safer, more reliable LVLMs, though cognition-based gains remain more challenging and require knowledge integration beyond perception enhancements.

Abstract

Large Vision-Language Models (LVLMs) demonstrate remarkable capabilities in multimodal tasks, but visual object hallucination remains a persistent issue. It refers to scenarios where models generate inaccurate visual object-related information based on the query input, potentially leading to misinformation and concerns about safety and reliability. Previous works focus on the evaluation and mitigation of visual hallucinations, but the underlying causes have not been comprehensively investigated. In this paper, we analyze each component of LLaVA-like LVLMs -- the large language model, the vision backbone, and the projector -- to identify potential sources of error and their impact. Based on our observations, we propose methods to mitigate hallucination for each problematic component. Additionally, we developed two hallucination benchmarks: QA-VisualGenome, which emphasizes attribute and relation hallucinations, and QA-FB15k, which focuses on cognition-based hallucinations.

A Comprehensive Analysis for Visual Object Hallucination in Large Vision-Language Models

TL;DR

This work performs a component-level analysis of visual object hallucination in LVLMs, decomposing sources of error into the LLM, vision backbone, and projector. It demonstrates that LLMs can faithfully describe content when provided accurate textual captions, while hallucinations primarily emanate from the vision encoder and projector. The authors propose targeted mitigations: fine-grained CLIP tuning and perception-based visual instruction tuning to strengthen the vision backbone, and contrastive alignment losses to improve projector alignment, alongside two fine-grained benchmarks (QA-VisualGenome and QA-FB15K) to evaluate perception and cognition hallucinations. The results show meaningful reductions in perception-based hallucinations and gains in cognition-based tasks, highlighting practical improvements for safer, more reliable LVLMs, though cognition-based gains remain more challenging and require knowledge integration beyond perception enhancements.

Abstract

Large Vision-Language Models (LVLMs) demonstrate remarkable capabilities in multimodal tasks, but visual object hallucination remains a persistent issue. It refers to scenarios where models generate inaccurate visual object-related information based on the query input, potentially leading to misinformation and concerns about safety and reliability. Previous works focus on the evaluation and mitigation of visual hallucinations, but the underlying causes have not been comprehensively investigated. In this paper, we analyze each component of LLaVA-like LVLMs -- the large language model, the vision backbone, and the projector -- to identify potential sources of error and their impact. Based on our observations, we propose methods to mitigate hallucination for each problematic component. Additionally, we developed two hallucination benchmarks: QA-VisualGenome, which emphasizes attribute and relation hallucinations, and QA-FB15k, which focuses on cognition-based hallucinations.
Paper Structure (30 sections, 14 equations, 3 figures, 12 tables)

This paper contains 30 sections, 14 equations, 3 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Hallucination Analysis
  3. Settings
  4. Language Decoder
  5. Conjecture 1. LLM in LVLM is able to generate faithful content when image captions are provided as input.
  6. Vision Encoder
  7. Conjecture 2. There are hallucinations in the perception process of the vision encoder.
  8. Projector
  9. Mitigating Object Hallucination Caused by Different Modules
  10. How to alleviate the hallucination caused by CLIP?
  11. How to reduce hallucination caused by the projector?
  12. Loss Objectives
  13. Integrated Alignment Loss
  14. Integrated Alignment Loss
  15. Separate Contrastive Alignment Loss
  16. ...and 15 more sections

Figures (3)

  • Figure 1: An overview of our paper. We first investigate the sources of hallucination from a component-level perspective within the LVLM architecture. Based on the identified causes, we then design targeted methods to mitigate hallucinations effectively.
  • Figure 2: Tuning CLIP with fine-grained data (left) and fine-grained perception-based instruction tuning (right).
  • Figure 3: The illustration of the hallucinated case for CLIP and LLaVA.