Exploring Audio Hallucination in Egocentric Video Understanding

1. Problem Statement

While Audio-Visual Large Language Models (AV-LLMs) have demonstrated impressive capabilities in generating multimodal descriptions, they suffer from a critical flaw: audio hallucination. This occurs when models infer sounds based on visual cues rather than actual auditory input.

This issue is particularly acute in egocentric (first-person) video, where:

• Visual inputs are often unstable, occluded, or blurry due to continuous camera movement.
• Audio serves as a crucial, reliable cue for understanding user activities and surroundings.
• Current models exhibit a strong visual bias, often generating sound descriptions that match what is seen (e.g., seeing a shovel implies a "digging sound") even when the audio track is silent or contains different sounds.

Prior research has focused heavily on visual hallucinations, leaving audio hallucinations in first-person settings largely underexplored.

2. Methodology

The authors propose a systematic framework to evaluate and analyze these hallucinations through a curated dataset and a targeted Question-Answering (Q/A) protocol.

A. Dataset Curation

• Source: 300 clips from the Ego4D corpus, covering 50 distinct visual scenarios.
• Pipeline (Fig. 2):
  1. Segmentation: Source audio is split into non-overlapping 10-second clips.
  2. Captioning: Qwen2.5 Omni generates preliminary audio captions.
  3. Stage 1: Sound-Source Match: LLaMA3-70B correlates audio captions with human-annotated narrations (time-aligned actions). It classifies sounds into:
     • Foreground Sounds: Directly caused by the user's actions (e.g., "mechanical sound" from "using a blender").
     • Background Sounds: Ambient environmental sounds independent of the user (e.g., wind, distant music).
  4. Stage 2: Q/A Generation: LLaMA3-70B generates 1,000 Q/A pairs, including:
     • Factual Q/As: Asking about present sounds.
     • Hallucination Q/As: Asking about plausible but absent sounds (e.g., asking for the source of a "hissing sound" in a scene where no frying occurs) to test if the model fabricates an answer based on visual context.

B. Evaluation Framework

• Models Evaluated: Four state-of-the-art AV-LLMs: ImageBind-LLM, VideoLLaMA2, Qwen2.5-Omni (7B), and MiniCPM (8B).
• Metrics: Three automatic scoring schemes were compared against human evaluation on a subset of 200 pairs:
  1. String Matching (Regex).
  2. Embedding Similarity (NV-Embed-v2).
  3. LLM-as-Judge (using GPT-4o).
• Result: The LLM-as-Judge approach showed the highest agreement with human evaluation and was used for the main experiments.

3. Key Contributions

1. First Taxonomy-Driven Evaluation: Introduced a grounded taxonomy specifically for egocentric videos, distinguishing between foreground action sounds and background ambient sounds.
2. Benchmark Creation: Curated a dataset of 300 videos and 1,000 targeted Q/A pairs designed to probe both factual audio understanding and susceptibility to visual-induced hallucinations.
3. Systematic Analysis: Demonstrated that even advanced models rely heavily on visual priors, failing to ground audio content accurately when visual and auditory cues conflict.

4. Key Results

The evaluation revealed significant limitations in current AV-LLMs:

• High Hallucination Rates:
  • Qwen2.5 Omni (the strongest model) achieved only 27.3% accuracy on foreground hallucination Q/As and 39.5% on background hallucination Q/As.
  • In contrast, factual Q/A accuracy was significantly higher (e.g., 56.2% for foreground and 63.4% for background), indicating models can identify what is there but struggle to confirm what is not.
• Visual Bias: Models consistently performed better on background sounds than foreground sounds. The authors argue this is because models exploit visual cues for background context but fail to distinguish specific user actions (foreground) from the visual scene when audio is missing or ambiguous.
• Qualitative Failure Modes:
  • Imprecise Grounding: Models fail to pinpoint the exact source of a real sound (e.g., guessing "knocking something over" instead of "closing drawers").
  • Cross-Modal Hallucination: Models confidently invent sound sources based on visual objects (e.g., claiming music is coming from a "handbag" or "speaker" when no music is playing).

5. Significance and Conclusion

• Reliability Gap: The study highlights that current AV-LLMs are unreliable for real-world multimodal scenarios where audio verification is critical, especially in first-person views where vision is often compromised.
• Need for Robust Evaluation: The paper argues that standard multimodal benchmarks are insufficient. A robust evaluation of hallucinations is essential for developing trustworthy AV-LLMs.
• Future Directions: The findings suggest that future models must be trained to better decouple visual priors from auditory evidence, ensuring that sound descriptions are strictly grounded in the audio track rather than inferred from the visual scene.

Project Page: https://cs20s030.github.io/EgoAVU/