permuto_sdf icon indicating copy to clipboard operation
permuto_sdf copied to clipboard

How to run custom dataset

Open JyVenom opened this issue 1 year ago • 13 comments

Not an issue, but how I did it.

In Colmap, perform the automatic reconstruction (untoggle dense since it is not needed)

Then, in the File pane, choose "Export model as text" and it will export a cameras.txt and images.txt file.

cameras.txt contains fx, fy, cx, and cy (It's the camera parameters. You can look up what the params mean for different camera types. For simple radial, since square pixels, fx = fy = params[0], cx = params[1], and cy = params[2])

images.txt contains tx, ty, tz, qw, qx, qy, and qz for each image. You must transform these 7 values to get the value in the world coordinate system. Apply -R^t * T (https://colmap.github.io/format.html and https://stackoverflow.com/questions/69210497/how-to-visualize-colmap-export-images-txt-in-blender).

Finally combine the new 7 values with the camera intrinsic to generate a poses_and_intrinsics.txt. Look at the file in the EasyPBR head dataset as an example for the order of variables. Make sure to toggle checkpoint saving in the config file.

When running the run_custom_dataset.py script, make sure to scale and translate as necessary.

Finally, to get your mesh, make a copy of the create_my_meshes.py script, delete all references to DataLoader's (they are created but not used), and run the new script (you will need to update the checkpoints and training scene paths files).

JyVenom avatar Jun 14 '23 09:06 JyVenom