Enhancing Visual Perception in Vision-Language Models with TWIN Dataset

This paper introduces a novel training dataset and task (TWIN) designed to improve the fine-grained visual perception capabilities of Vision-Language Models (VLMs). The core idea is to train VLMs to distinguish between visually similar images of the same object, forcing them to attend to subtle visual details. The paper demonstrates significant improvements on fine-grained recognition tasks and introduces a new benchmark (FGVQA) to quantify these gains. The work addresses a key limitation of current VLMs and provides a practical contribution in the form of a new dataset and training methodology.