Evolution of Accuracy and Visual-Cognitive Errors in a Decade of Vision-Language AI Models
Vision language models (VLMs) have made remarkable progress in visual reasoning during the last decade. Most evaluations have used simple scenes (MS-COCO) that do not showcase complex human interactions or behaviors, only a handful of non-curated human descriptions as a benchmark, and have not focused on understanding visual-cognitive errors.
A decade of vision-language AI models shows progress in visual reasoning, but evaluations have relied on simple scenes and limited benchmarks, lacking focus on visual-cognitive errors.
The paper suggests that current VLM evaluations may underestimate errors in complex human interactions, indicating a need for more challenging benchmarks.
The finding implies that VLM performance in real-world applications involving human behavior may be overestimated, potentially affecting deployment in areas like autonomous driving or surveillance.
Improved evaluation could reduce deployment risks for VLMs in high-stakes applications, increasing trust and adoption.
Future work may develop benchmarks targeting visual-cognitive errors in complex scenes, leading to more robust VLMs.