ORSet
收藏资源简介:
ORSet是由华中科技大学团队构建的首个全向参考多目标跟踪数据集,基于JackRabbot数据集扩展而来。该数据集包含27个多样化全向场景、848条涵盖外观/动作/空间关系的语言描述,以及3401个带有边界框和时空轨迹的标注对象,数据总量达17个训练场景和10个测试场景。通过三阶段标注流程(关键帧选择-GPT-4o描述生成-人工校验对齐)确保数据质量,特别关注全向相机特有的长时程语义理解。该数据集旨在解决传统受限视场相机导致的跟踪碎片化问题,推动360°视觉-语言多模态对齐、零样本泛化和时序 grounding 等研究方向。
ORSet is the first omnidirectional reference multi-object tracking dataset constructed by the team from Huazhong University of Science and Technology, which is extended from the JackRabbot dataset. This dataset contains 27 diverse omnidirectional scenarios, 848 language descriptions covering appearance, action, and spatial relationships, as well as 3401 annotated objects with bounding boxes and spatiotemporal trajectories. In total, it comprises 17 training scenarios and 10 test scenarios. Data quality is ensured via a three-stage annotation workflow: keyframe selection, GPT-4o-based description generation, and manual verification and alignment, with special emphasis on the long-term semantic understanding unique to omnidirectional cameras. This dataset aims to address the tracking fragmentation problem caused by traditional limited field-of-view cameras, and advance research in fields including 360° vision-language multimodal alignment, zero-shot generalization, and temporal grounding.
ORMOT数据集概述
数据集基本信息
- 数据集名称:ORMOT (Omnidirectional Referring Multi-Object Tracking)
- 核心内容:一个用于全向参考多目标跟踪的数据集与框架。
- 关联项目:与CRMOT项目类似(https://github.com/chen-si-jia/CRMOT)。
数据集特点与定义
- 任务定义:全向参考多目标跟踪。
- 技术特点:利用全向相机宽广的视野,不仅提供空间优势,还通过提供“扩展的时间上下文”来延长跟踪持续时间,使模型能够正确理解长视野语言描述并准确跟踪目标。
- 对比任务:与传统参考多目标跟踪相比,传统相机视野有限,使得现有常见RMOT模型更难以理解长视野语言描述并执行精确跟踪。
数据获取与状态
- 当前状态:论文若被接受,作者将在一个月内完全开源ORSet数据集和ORTrack框架(包括其代码和模型权重)。
- 开源计划:遵循其CRMOT项目的模式。
论文与引用
- 论文标题:ORMOT: A Dataset and Framework for Omnidirectional Referring Multi-Object Tracking
- 预印本地址:https://arxiv.org/pdf/2603.05384
- 作者:Sijia Chen, Zihan Zhou, Yanqiu Yu, En Yu, Wenbing Tao
- 机构:华中科技大学
- 发表信息:arXiv preprint arXiv:2603.05384, 2026
- 引用格式:
@article{chen2026ormot, title={ORMOT: A Dataset and Framework for Omnidirectional Referring Multi-Object Tracking}, author={Chen, Sijia and Zhou, Zihan and Yu, Yanqiu and Yu, En and Tao, Wenbing}, journal={arXiv preprint arXiv:2603.05384}, year={2026} }

- 1ORMOT: A Dataset and Framework for Omnidirectional Referring Multi-Object Tracking华中科技大学·多谱信息智能处理技术国家重点实验室 · 2026年



