VIABench: A New Benchmark for Evaluating AI in Visual Impairment Assistance

Researchers have introduced VIABench, a comprehensive video benchmark designed to evaluate Multimodal Large Language Models (MLLMs) in assisting visually impaired individuals. VIABench focuses on three core tasks: Proactive Reminder, Visual Question Answering (VQA), and Vision-Guided Interaction. The benchmark utilizes first-person videos recorded by visually impaired individuals to assess the models' capabilities in real-world scenarios. Initial experiments indicate that current MLLMs struggle, particularly with the Proactive Reminder task, highlighting the need for further development in this area. The research team aims to drive future advancements in AI to improve navigation and interaction experiences for the visually impaired. The code and data for VIABench are available at https://github.com/MCG-NJU/VIABench.