OVRSISBench
收藏资源简介:
OVRSISBench是一个针对开放词汇遥感图像分割(OVRSIS)任务的统一基准,基于广泛使用的遥感分割数据集构建。该数据集包含8个代表性的遥感数据集,包括DLRSD、iSAID、Potsdam、Vaihingen、UAVid、UDD5、LoveDA和VDD,涵盖了城市布局、农业区域和高分辨率航空影像等多种场景。OVRSISBench采用开放词汇范式,要求模型根据文本描述泛化到未见过的类别。通过重新制定这些数据集,OVRSISBench保留了传统遥感任务的领域特定挑战,同时结合了开放词汇分割的灵活性和泛化需求。该基准为开放词汇分割模型在真实遥感场景下的公平、一致和可扩展评估提供了便利。
OVRSISBench is a unified benchmark for the Open-Vocabulary Remote Sensing Image Segmentation (OVRSIS) task, constructed based on widely adopted remote sensing segmentation datasets. This benchmark comprises eight representative remote sensing datasets, namely DLRSD, iSAID, Potsdam, Vaihingen, UAVid, UDD5, LoveDA and VDD, covering diverse scenarios such as urban layouts, agricultural regions and high-resolution aerial imagery. OVRSISBench adopts the open-vocabulary paradigm, which requires models to generalize to unseen categories based on textual descriptions. By reformulating these datasets, OVRSISBench retains the domain-specific challenges inherent in traditional remote sensing tasks, while integrating the flexibility and generalization demands of open-vocabulary segmentation. This benchmark facilitates fair, consistent and scalable evaluation of open-vocabulary segmentation models in real-world remote sensing scenarios.

- 1通过中国科学院大学,西北工业大学,中国人工智能研究院 · 2025年



