SceneParser/SceneParser-bench

SceneParser-Bench是一个用于层次化场景解析的大规模基准数据集，基于EgoObjects构建。它包含11万张训练图像和5千张验证图像，图像之间没有重叠，并标注了明确的层次结构，将局部化对象、语义部件和功能性链接起来。数据集总计包含174万有效的对象-部件-功能性链实例。数据格式为JSONL，每个样本将RGB图像链接到一个结构化的“场景 -> 对象 -> 部件 -> 功能性”层次结构，用于训练和评估SceneParser模型。SceneParser-Bench is constructed based on EgoObjects as a large-scale benchmark for hierarchical scene parsing. It contains 110K training images and 5K validation images with no image overlap, annotated with explicit hierarchies linking localized objects, semantic parts, and affordances. In total, SceneParser-Bench contains 1.74M valid object-part-affordance chain instances. Each sample links an RGB image to a structured `scene -> object -> part -> affordance` hierarchy for training and evaluating SceneParser.