merlerm/clevr_hard

CLEVR-Hard是一个视觉问答数据集，使用CLEVR基元（立方体、球体、圆柱体）渲染生成。每个问题有两种变体：原始场景渲染和干扰项变体（非保护属性随机化）。数据集包含50个验证场景，每个场景有6个对象和10个问题。活动属性包括颜色、形状、材质、大小和纹理。图像尺寸为480x320，渲染样本数为64。数据集字段包括ID、场景ID、变体类型、图像、问题、答案、问题类型以及一些JSON字符串字段（protected、forbidden、identifying_filters）。

CLEVR-Hard is a visual question answering dataset rendered with CLEVR primitives (cubes, spheres, cylinders). Each question has two variants: original scene rendering and distractors variant (non-protected attributes randomized). The dataset contains 50 validation scenes, with 6 objects and 10 questions per scene. Active attributes include color, shape, material, size, and texture. Image dimensions are 480x320 with 64 render samples. Fields include id, scene_id, variant type, image, question, answer, question_type, and some JSON string fields (protected, forbidden, identifying_filters).