Dataset for Street-level Surface Reconstrucion

The dataset offers a comprehensive set of resources, including aligned RGB images and LiDAR point cloud frames, frame poses, calibration parameters, and ground truth point clouds, for street-level surface reconstruction. The dataset is collected by a handheld camera-LIDAR-IMU integrated device. The device is equipped with an IMU, a monocular camera, and a Velodyne VLP-16 LiDAR. We collect 5 sequences, namely real_1, real_2, real_3, real_4, and real_5, respectively. The first four sequences are captured along line-shaped streets with lengths of about 230 m, 235 m, 225 m, and 410 m, respectively. The last sequence is collected in a small park with an area of about 370 m * 100 m, and the data collection trajectory is a ring with a length of about 870 m. We perform strict spatiotemporal calibration to determine the camera's internal parameters as well as external parameters among the camera, LiDAR, and IMU. We use a robust LiDAR-inertial odometry, FAST-LIO2, to obtain the initial poses of each data frame. A NavVis VLX 3 scanner is used to capture the ground truth point cloud of each sequence with millimeter-level accuracy. To obtain the ground truth poses of each frame, we first compute a transformation between the accumulated point cloud generated by Fast-LIO2 and the ground truth point cloud, then apply the transformation to the initial poses, and finally refine the poses by aligning each point cloud frame with the ground truth point cloud using the iterative closest point (ICP) algorithm.