Multi-Object Hallucination in Vision-Language Models

TL;DR

ROPE is introduced, an automated evaluation protocol that considers the distribution of object classes within a single image during testing and uses visual referring prompts to eliminate ambiguity, to enable LVLMs to recognize and reason about multiple objects that often occur in realistic visual scenes.

Abstract

Large vision language models (LVLMs) often suffer from object hallucination, producing objects not present in the given images. While current benchmarks for object hallucination primarily concentrate on the presence of a single object class rather than individual entities, this work systematically investigates multi-object hallucination, examining how models misperceive (e.g., invent nonexistent objects or become distracted) when tasked with focusing on multiple objects simultaneously. We introduce Recognition-based Object Probing Evaluation (ROPE), an automated evaluation protocol that considers the distribution of object classes within a single image during testing and uses visual referring prompts to eliminate ambiguity. With comprehensive empirical studies and analysis of potential factors leading to multi-object hallucination, we found that (1). LVLMs suffer more hallucinations when focusing on multiple objects compared to a single object. (2). The tested object class distribution affects hallucination behaviors, indicating that LVLMs may follow shortcuts and spurious correlations. (3). Hallucinatory behaviors are influenced by data-specific factors, salience and frequency, and model intrinsic behaviors. We hope to enable LVLMs to recognize and reason about multiple objects that often occur in realistic visual scenes, provide insights, and quantify our progress towards mitigating the issues.

Figures (8)

• Figure 1: A case study that compares our Recognition-based Object Probing Evaluation (ROPE) benchmark with existing benchmarks for object hallucination in GPT-4V. ROPE offers an automated evaluation protocol with controlled output formatting and uses visual prompts to distinctly ground to objects, thus mitigating referential ambiguity. Unlike binary inquiries relying solely on textual descriptions, ROPE challenges the model to identify multiple objects concurrently. We observe that, while GPT-4V can identify the whisk to the left of a knife when prompted about it, the model hallucinates a "fork" when directly tasked to recognize multiple objects.
• Figure 2: A heterogeneous ROPE sample tested with Deafult multi-object query, where each of the 5 objects belongs to different object classes. We label the output class as either correct or hallucinated.
• Figure 3: A homogeneous ROPE sample, where the 5 objects belong to the same object class, and a corresponding adversarial ROPE sample, where the last object belongs to a different object class.
• Figure 4: The performance of the LLaVA on the adversarial split, organized by the query sequence of AAAAB and BAAAA, reveals significant vulnerabilities as the model's accuracy dramatically declines for object 5 in AAAAB. SO stands for single-object probing and TF stands for teacher-forcing probing.
• Figure 5: A comparison of the distribution of hallucinatory versus non-hallucinatory object classes in LLaVA-13B, across the unseen split under student forcing.