This is the Heigold Facade, a historical landmark in Louisville, KY. It was imaged by Eric Stemen in (April?) 2026 using a 3DMakerPro Raven handheld SLAM Lidar Scanner. He graciously shared his dataset online for people to "play around with these files yourself". I downloaded the files and applied some custom offline processing I'm working to develop for use with a Raven scanner I ordered and hope to receive soon.
The processing steps are as follows:
(1) Run Raystudio to compute the point cloud.
(2) In Raystudio run Gaussian Splatting just until the point it exports the full fisheye images.
(3) Run a python script to match left & right camera images and then move two of each sampled at 2 frames per second to a subdirectory with separate left and right camera subdirectories.
(4) Run Exiftool to label the images from the left and right cameras with separate camera make and model so they are identified as coming from two different separate cameras.
(5) Using Agisoft Metashape Standard, load the images and set the cameras type to "Equisolid Fisheye" and turn on Full Rolling Shutter Correction.
(6) Hand draw lens (to remove the dark round image dark borders) and operator masks. (This should, instead by done by AI like SAM3).
(7) Align the images.
(8) Export the cameras in Colmap format (with the images and masks) Do it separately for the right and left cameras so they are in separate directories.
(9) Export the computed camera lens model for each lens in Agisoft XML format.
(10) Run a custom Python script to convert each fisheye image to five pinhole images (cubemap faces). Do it separately for each lens using the Metashape lens parameters. (https://github.com/alexmgee/Fisheye-to-Cubemap)
(11) Load all of the pinhole images and masks into Metashape and set the image groups to "camera stations".
(12) Align the images.
(13) Export the cameras in Colmap format with the images masks and point cloud.
(14) Train a Gaussian Splat in Brush.
(15) Edit the Gaussian Splat in Supersplat.
Eventually I hope to automate 9-13 to avoid a second image alignment.