Gaussian Splatting on the Move - Smartphone Dataset

Casual smartphone 3D scans using three devices: iPhone 15 Pro, Samsung Galaxy S20 FE and Google Pixel 5. The raw data (prefixed "spectacular-rec-") has been captured with the Spectacular Rec applications for Android and iOS, and contains time-synchronized video and IMU data. The extras file also contains AprilGrid calibration sequences, as well as pre-computed calibration results, for the Android devices.

The data is captured in non-ideal lighting conditions and has a moderate amount of motion blur and rolling shutter artefacts. The included metadata also contains the exposure times and the (Android) rolling shutter readout times, as well as the built-in calibration data, as reported by the devices.

The dataset also contains three different processed variants (prefixed with "colmap-"), which are directly trainable with Nerfstudio. In the processed variants, suitable minimally blurry video frames have been selected as key frames and their poses have been registered with COLMAP. In addition, the local linear and angular velocities of each key frame has been estimated using VIO with Spectacular AI Mapping Tools. The "calib-intrinsics" and "orig-intrinsics" variants include manually calibrated and built-in intrinsics, respectively. They depend on the third variant with symbolic links. The third variant has COLMAP-estimated intrinsics, which are relatively inaccurate for the Android data with high levels rolling shutter deformation.

本数据集采用三款设备采集日常休闲场景下的智能手机3D扫描数据，分别为iPhone 15 Pro、三星Galaxy S20 FE以及谷歌Pixel 5。原始数据（文件名前缀为"spectacular-rec-"）通过适用于Android与iOS平台的Spectacular Rec应用采集，包含时间同步的视频数据与惯性测量单元（Inertial Measurement Unit，IMU）数据。额外文件还包含针对Android设备的April格网（AprilGrid）标定序列与预计算标定结果。

本数据集采集于非理想光照环境，存在中等程度的运动模糊与卷帘快门伪影（rolling shutter artefacts）。附带的元数据包含曝光时长、（Android设备的）卷帘快门读出时长，以及设备上报的内置标定数据。

本数据集还包含三种以"colmap-"为前缀的预处理变体，可直接用于Nerfstudio的模型训练。在预处理变体中，已筛选出模糊程度较低的视频帧作为关键帧，并通过COLMAP完成其位姿配准。此外，已借助Spectacular AI Mapping Tools结合视觉惯性里程计（Visual-Inertial Odometry，VIO）估算出每个关键帧的局部线速度与角速度。其中"calib-intrinsics"与"orig-intrinsics"两种变体分别对应手动标定内参与设备内置内参，二者均依赖于带有符号链接的第三种变体。第三种变体采用COLMAP估算的内参，对于存在严重卷帘快门变形的Android设备数据而言，其标定精度相对较低。