Colmap with load_llff function

[massyzs]

I found that if "args.dataset_type == 'llff' ", the code will select test images from the data. I'm wondering how could you make sure the depth points are the same if generated without test views?

For example, if you have 3 images as training view. To get depths estimated by three view, you cannot add the fourth (test view) into colmap. But in that case, you cannot generate the consistent camera pose of the fourth image and you cannot evaluate it.