OpenLex3D
收藏资源简介:
OpenLex3D是一个专为评估3D开放词汇场景表示的新基准。该数据集由Replica、ScanNet++和HM3D三个数据集中的23个场景构成,提供了全新的标签注释,捕捉了现实世界中的语言变异性,引入了同义词物体类别和额外的细微描述。数据集通过引入开放集3D语义分割任务和基于文本查询的对象检索任务,来评估特征精度、分割效果和下游能力。
OpenLex3D is a novel benchmark specifically designed for evaluating 3D open-vocabulary scene representations. It consists of 23 scenes curated from three existing datasets: Replica, ScanNet++, and HM3D. The dataset provides newly developed label annotations that capture real-world linguistic variability, introducing synonymous object categories and additional fine-grained descriptive details. To evaluate feature accuracy, segmentation performance and downstream capabilities, the benchmark incorporates two tasks: open-set 3D semantic segmentation and text-query-based object retrieval.




