OmniScene
收藏资源简介:
OmniScene是一个用于深度估计任务的光线级逼真合成数据集,由NVIDIA Isaac Sim渲染生成。该数据集专为训练X-Lens模型而设计,通过一个6相机阵列采集每个场景,包括4个鱼眼相机(提供全向视角)和2个针孔相机(提供前向和后向视角),所有相机分辨率均为1920×1200。每个样本同时包含针孔、鱼眼及异构混合视图,并附带度量级真实深度标签。数据集按场景打包,每个场景为一个.tar文件,包含训练集、验证集和测试集分割。每个场景包含以下内容:RGB图像(8位JPG格式)、深度图(16位PNG格式,需除以256转换为米制深度)、每相机的静态有效镜头掩码、天空掩码(当场景包含天空时)以及每帧的相机参数(包含所有6个视角的内参、外参和完整标定信息)。数据集规模在100B到1T之间,适用于深度估计、多视角视觉、鱼眼图像处理等任务,特别适合研究异构相机系统的深度感知。数据集采用CC BY-NC 4.0许可,仅限非商业研究使用。
OmniScene is a photorealistic synthetic dataset for depth estimation tasks, rendered using NVIDIA Isaac Sim. It is designed for training the X-Lens model, capturing each scene with a 6-camera array consisting of 4 fisheye cameras (providing omnidirectional views) and 2 pinhole cameras (providing forward and backward views), all with a resolution of 1920×1200. Each sample includes pinhole, fisheye, and heterogeneous hybrid views, along with metric-accurate ground truth depth labels. The dataset is packaged by scene, with each scene as a .tar file containing splits for training, validation, and testing. Each scene includes: RGB images (8-bit JPG format), depth maps (16-bit PNG format, requiring division by 256 to convert to metric depth), per-camera static valid lens masks, sky masks (when the scene includes sky), and per-frame camera parameters (including intrinsic, extrinsic, and full calibration information for all 6 views). The dataset size ranges from 100B to 1T, suitable for tasks such as depth estimation, multi-view vision, and fisheye image processing, particularly for research on depth perception in heterogeneous camera systems. It is licensed under CC BY-NC 4.0 for non-commercial research use only.
数据集概览
OmniScene 是一个用于深度估计的合成数据集,专为训练 X-Lens 模型而设计。
- 渲染工具: NVIDIA Isaac Sim
- 采集设备: 6 摄像头阵列(4 个鱼眼 + 2 个针孔)
- 标注: 包含度量真实深度(metric ground-truth depth)
- 分辨率: 1920×1200
- 许可协议: CC BY-NC 4.0(仅限非商业研究用途)
数据规模与划分
- 文件大小: 100B < n < 1T
- 划分: 训练集(
train/)、验证集(valid/)、测试集(test/),每个场景打包为单个.tar文件
文件结构
train/<scene>.tar valid/<scene>.tar test/<scene>.tar texture/ # 共享的鱼眼光线方向查找表(LUT)
每个 .tar 解压后的目录结构为:
<scene>/ ├── rgb/ CAM_/<frame>.jpg # 8-bit RGB 图像 ├── depth/ CAM_/<frame>.png # 16-bit 度量深度图 ├── mask/ CAM_*_mask.png # 每个相机的静态有效镜头遮罩 ├── sky_mask/ _meta.json + 帧级天空遮罩 └── common/ <frame>.npy # 所有6个视角的相机参数
摄像头阵列
| 视角 | 模型 | 说明 |
|---|---|---|
| CAM_A, CAM_B, CAM_C, CAM_D | 鱼眼(全向,Mei / unified-sphere: xi + radtan) | 4 个侧面摄像头 |
| CAM_Front, CAM_Back | 针孔(透视,3×3 内参 K) | 前后摄像头 |
深度解码
- 深度图为 16-bit PNG,真实深度(米)= 像素值 / 256
- 像素值为 0 表示无效区域或天空(需配合
sky_mask使用) - 加载方法:
cv2.imread(path, cv2.IMREAD_UNCHANGED)确保读取16位数据
使用示例
python from huggingface_hub import hf_hub_download import tarfile
下载单个场景
p = hf_hub_download("henryzhou998/OmniScene", "test/<scene>.tar", repo_type="dataset") tarfile.open(p).extractall("omniscene/")
下载整个测试集划分
from huggingface_hub import snapshot_download snapshot_download("henryzhou998/OmniScene", repo_type="dataset", allow_patterns="test/*", local_dir="omniscene_tars")
许可证
- CC BY-NC 4.0(非商业研究使用)




