APRS benchmark
收藏资源简介:
APRS基准数据集是由复旦大学研究团队构建的,旨在支持主动全景指代分割任务的大规模评估资源。该数据集包含来自360-Indoor、PANDORA和SUN360的4,971个独特全景场景,共计7,420个标注样本,涵盖了多样化的室内外环境,并设计了包含空间关系推理的复杂语言指令。数据集的创建过程采用了人机协同的标注流程,首先由专业标注员标注显著目标并设置初始视角,随后结合SAM-3等基础模型生成初始掩码并人工 refinement,同时利用VLM进行语义交叉验证以确保可靠性。该数据集主要应用于具身人工智能和计算机视觉领域,旨在解决智能体在连续360度环境中进行主动感知、跨视角空间推理以及目标分割的综合性挑战,推动从被动静态图像处理到主动多视角交互的范式转变。
The APRS Benchmark Dataset was constructed by a research team from Fudan University, serving as a large-scale evaluation resource to support the active panoramic referring segmentation task. This dataset includes 4,971 unique panoramic scenes sourced from 360-Indoor, PANDORA, and SUN360, with a total of 7,420 annotated samples. It covers diverse indoor and outdoor environments, and features complex language instructions that incorporate spatial relation reasoning. The dataset was developed through a human-machine collaborative annotation workflow: first, professional annotators label salient objects and set initial viewing perspectives; subsequently, foundational models such as SAM-3 are employed to generate initial masks, followed by manual refinement, and VLMs are utilized for semantic cross-validation to ensure annotation reliability. This dataset is primarily applied in the fields of embodied artificial intelligence and computer vision, aiming to address the comprehensive challenges faced by AI agents when conducting active perception, cross-view spatial reasoning, and object segmentation in continuous 360-degree environments, and promote the paradigm shift from passive static image processing to active multi-view interaction.
数据集概述
任务定义
Active Panoramic Referring Segmentation (APRS):一个全新的任务,要求智能体在连续的 360° 环境中主动探索,通过调整相机视角(θ, φ)来寻找并分割用户指令指定的目标对象。该任务将传统的被动式指代分割(基于单张固定图像)转变为主动式(探索整个 360° 球面空间)。
数据集规模
- 4,971 个独特的 360° 场景
- 7,420 个带有专家标注的样本
场景与标注类型
数据集涵盖多样的室内/室外场景,并包含四种类型的空间指代表达式:
- EGO(自我中心):相对于智能体朝向的指令,例如“向右转找到沙发。”
- UNIQ(独特属性):基于显著视觉属性的指令,例如“房间里黄色的落地灯。”
- ALLO(异我中心):相对于锚定物体的指令,例如“沙发对面的椅子。”
- MULTIHOP(多跳):多步骤组合推理的指令,例如“转过身,找到桌子旁边的椅子。”
评价协议
数据集提供了全面的评价协议,主要指标包括:
- SR(Success Rate):成功率
- AS(Average Steps):平均步骤数
- SPL(Success weighted by Path Length):路径长度加权成功率
- mIoU(mean IoU):平均交并比(分割质量)
基线对比
在 APRS 基准上,PanoSeeker 在与静态方法、启发式方法和基于 VLM 的智能体(包括 GPT-5.2 和 Gemini-3)的对比中,在成功率、搜索效率和分割质量上均取得了领先结果。

- 1Seek to Segment: Active Perception for Panoramic Referring Segmentation复旦大学 · 2026年



