Really need the code files for stitching point clouds into a panoramic map.

[AlcApollo]

Excellent work—it’s very thoughtful of you to include the data map.

I’d like to try reproducing the results of this paper on other datasets, and it’s clear that the accuracy of map.ply will affect the quality of Gaussian reconstruction. Could you please provide the code for the point cloud stitching part?

Looking forward to your reply. Thank U.

[zang09]

Hi @AlcApollo, and sorry for the late reply.

There is no special trick involved in how we create the map. As long as you have the LiDAR poses, either from ground truth or a LiDAR SLAM system, you can simply transform each point cloud into the world coordinate system, merge them, and save the resulting point cloud.

One important detail is that our system assumes a rough initial rotation between the LiDAR and camera coordinate systems. In particular, we use dataset-specific camera yaw values such as:

DATASET_CAMERA_YAW_DEGREES = {
    "kitti-360": {2: 90.0, 3: -90.0},
    "kitti": {},
    "waymo": {1: 45.0, 2: -45.0, 3: 90.0, 4: -90.0},
    "fast-livo2": {},
}

You can find the relevant implementation in pose_utils.py.

Please check whether this initial rotation assumption is appropriate for your dataset. If the reconstruction result is poor, the LiDAR and camera coordinate systems may not be aligned correctly in the world coordinate frame.

Best regards,
Haebeom