FudanCVL/APRS_dataset
收藏资源简介:
APRS(主动全景参考分割)是一个用于360°全景环境中主动感知的大规模基准数据集。与处理静态图像的被动参考分割不同,APRS要求智能体通过调整观察方向来主动探索连续全景场景,以根据自然语言指令寻找和分割目标。数据集包含7,420个样本,覆盖4,971个多样化的全景场景,包括室内和室外环境。它提供四种类型的空间参考表达:以自我为中心(第一人称方向参考,如“向左看以找到...”)、独特属性(对象特征,如“红色沙发”)、以他者为中心(第三人称空间关系,如“靠近窗户的椅子”)和多跳(复杂关系推理,如“向左看找到床,然后找到旁边桌子上的台灯”)。
APRS (Active Panoramic Referring Segmentation) is a large-scale benchmark dataset for active perception in 360° panoramic environments. Unlike passive referring segmentation that processes static images, APRS requires agents to actively explore continuous panoramic scenes by adjusting viewing directions to seek and segment targets based on natural language instructions. The dataset includes 7,420 samples across 4,971 diverse panoramic scenes, covering both indoor and outdoor environments. It features four types of spatial referring expressions: Egocentric (first-person directional references, e.g., look left to find...), Unique-Attribute (distinctive object features, e.g., the red sofa), Allocentric (third-person spatial relations, e.g., the chair near the window), and Multi-hop (complex relational reasoning, e.g., look to the left to find the bed, then find the lamp on the table next to it).




