VisionTrap: Unanswerable Questions On Visual Data

TL;DR

This paper introduces VisionTrap, a dataset designed to evaluate whether visual question answering models can recognize unanswerable questions across diverse, unrealistically generated images, highlighting the need for models to abstain when appropriate.

Contribution

The paper presents a new dataset, VisionTrap, with unanswerable questions across various image types to test model recognition of limitations in VQA tasks.

Findings

Models often attempt to answer unanswerable questions.

Inclusion of unanswerable questions is crucial for comprehensive VQA evaluation.

VisionTrap reveals models' tendency to generate incorrect answers when unsure.

Abstract

Visual Question Answering (VQA) has been a widely studied topic, with extensive research focusing on how VLMs respond to answerable questions based on real-world images. However, there has been limited exploration of how these models handle unanswerable questions, particularly in cases where they should abstain from providing a response. This research investigates VQA performance on unrealistically generated images or asking unanswerable questions, assessing whether models recognize the limitations of their knowledge or attempt to generate incorrect answers. We introduced a dataset, VisionTrap, comprising three categories of unanswerable questions across diverse image types: (1) hybrid entities that fuse objects and animals, (2) objects depicted in unconventional or impossible scenarios, and (3) fictional or non-existent figures. The questions posed are logically structured yet inherently unanswerable, testing whether models can correctly recognize their limitations. Our findings highlight the importance of incorporating such questions into VQA benchmarks to evaluate whether models tend to answer, even when they should abstain.