MIRAGE
收藏资源简介:
MIRAGE是一个多模态基准数据集,旨在评估模型在计数(对象属性识别)、关系(空间关系推理)和计数与关系组合任务上的能力。数据集由1710个问题组成,涵盖广泛的视觉多样性和推理挑战,包括各种颜色、形状、大小和纹理的对象,以及从简单对到深层嵌套层次的空间关系。MIRAGE通过将计数和关系任务结合,测试模型在复杂场景中同时进行对象身份识别、数量统计和空间配置推理的能力。数据集的设计强调了视觉认知的组合性质,并针对表面级识别和关系理解之间的差距。
MIRAGE is a multimodal benchmark dataset designed to evaluate model capabilities in counting (object attribute recognition), relational (spatial relation reasoning), and combined counting and relational tasks. It consists of 1710 questions, covering a wide range of visual diversity and reasoning challenges, including objects with various colors, shapes, sizes and textures, as well as spatial relations ranging from simple pairs to deeply nested hierarchies. By integrating counting and relational tasks, MIRAGE tests models' ability to simultaneously perform object identity recognition, quantity statistics and spatial configuration reasoning in complex scenarios. The dataset's design emphasizes the compositional nature of visual cognition and targets the gap between surface-level recognition and relational understanding.

- 1MIRAGE: A Multi-modal Benchmark for Spatial Perception, Reasoning, and Intelligence独立研究者、清华大学、shopee、阿里巴巴国际数字商务、布里斯托尔大学 · 2025年



