MiaoMiaoYang/SCALAR-VG

SCALAR_VG数据集是一个大规模的多模态图像数据集，通过整合和扩展多个开源图像数据集构建而成，旨在支持大规模的多维场景感知理解训练。该数据集包含了大约24万张图像，并提供了全面、层次化、多维的注释，这些注释包括了几何描述符（边界框、关键点、分割多边形）、语义标识符（对象检测类别、参照性标题）和关系元数据（空间定位坐标、对象间关系图）。这些注释使得该数据集能够支持对低级视觉模式与高级上下文推理之间的整体场景感知理解。

The SCALAR_VG dataset is a large-scale multi-modal image dataset constructed by integrating and extending multiple open-source image datasets, aiming to support training for large-scale multi-dimensional scene-aware understanding. The dataset contains about 240,000 images, each annotated comprehensively, hierarchically, and multidimensionally. These annotations include geometric descriptors (bounding boxes, keypoints, segmentation polygons), semantic identifiers (object detection classes, referential captions), and relational metadata (spatial grounding coordinates, inter-object relationship graphs). These annotations enable the dataset to support holistic scene-aware understanding bridging low-level visual patterns with high-level contextual reasoning.