TJScenes
收藏资源简介:
TJScenes是由同济大学构建的用于高分辨率三维占据预测的专用数据集,旨在解决自动驾驶与机器人导航中细粒度场景理解的需求。该数据集包含27911个标注样本,采集自动态校园环境,每个样本提供六摄像头全景图像与0.1米分辨率的3D语义占据标注,数据来源于Livox Mid-360激光雷达扫描,覆盖±20米平面范围与-2米至4.4米高度。数据集创建通过定制移动机器人平台同步采集多传感器数据,并经过精细标注流程完成。其核心应用于评估高体素分辨率下的几何感知模型性能,特别关注人行道与非道路场景,弥补现有道路中心数据集的不足,推动机器人导航中的细粒度环境建模。
TJScenes is a dedicated dataset developed by Tongji University for high-resolution 3D occupancy prediction, designed to address the demand for fine-grained scene understanding in autonomous driving and robotic navigation. It comprises 27,911 annotated samples collected from dynamic campus environments. Each sample includes six-camera panoramic images and 3D semantic occupancy annotations with a 0.1-meter resolution. The data is captured via Livox Mid-360 LiDAR scans, covering a planar range of ±20 meters and a height span from -2 meters to 4.4 meters. The dataset is constructed by synchronously collecting multi-sensor data using a custom mobile robotic platform, followed by a rigorous annotation workflow. Its core application is to evaluate the performance of geometric perception models under high voxel resolution, with specific emphasis on sidewalk and off-road scenarios. It fills the gap of existing road-centric datasets and advances fine-grained environmental modeling for robotic navigation.
数据集名称
TJScenes
数据集简介
TJScenes 是一个全景六摄像头占用数据集,用于高分辨率三维几何感知评估。其标注具有0.1米的极高空间分辨率。
数据集用途
- 用于评估高分辨率三维占用预测任务,特别是面向自动驾驶和机器人导航的场景理解。
- 旨在结合所提出的 GaussianSeed 框架,推动高分辨率三维占用预测的效率与质量前沿。
数据集规模与分辨率
- 标注空间分辨率:0.1米。
- 传感器配置:全景六摄像头。
相关论文
- 论文标题:GaussianSeed: Hierarchical Gaussian Seeding for High-Resolution 3D Occupancy Prediction
- 作者:Xinzhuo Li, Xianghui Pan, Jiayuan Du, Wei Wei, Liuyi Wang, Chengju Liu, Qijun Chen
- 提交时间:2026年7月22日
- 论文链接:https://arxiv.org/abs/2607.20071v1

- 1GaussianSeed: Hierarchical Gaussian Seeding for High-Resolution 3D Occupancy Prediction同济大学 · 2026年



