Evolution of Accuracy and Visual-Cognitive Errors in a Decade of Vision-Language AI Models

Over the past decade, vision-language AI models have made significant strides in accurately describing scenes, often matching human performance on benchmarks like MS-COCO. However, these models struggle when scenes become complex, especially with multi-agent social dynamics or in less ideal lighting conditions. This limitation highlights a crucial gap in how these AIs interpret real-world scenarios compared to idealized test conditions. Understanding these visual-cognitive errors is essential for advancing AI that can handle the complexities of human environments.