How does split_train_test function work at all?

[dsvilarkovic]

Hello,
I just looked into your function that you use for initializing img_files to be used for spllitting data for Mast3r inference

#utils/sfm_utils.py
def split_train_test(image_files, llffhold=8, n_views=None, verbose=True):
    test_idx  = np.linspace(1, len(image_files) - 2, num=12, dtype=int)
    train_idx = [i for i in range(len(image_files)) if i not in test_idx]

    sparse_idx = np.linspace(0, len(train_idx) - 1, num=n_views, dtype=int)
    train_idx = [train_idx[i] for i in sparse_idx]

    if verbose:
        print(">> Spliting Train-Test Set: ")
        # print(" - sparse_idx:         ", sparse_idx)
        print(" - train_set_indices:  ", train_idx)
        print(" - test_set_indices:   ", test_idx)
    train_img_files = [image_files[i] for i in train_idx]
    test_img_files = [image_files[i] for i in test_idx]

    return train_img_files, test_img_files

From what I see here, there is no splitting according to LLFF logic, and this code does not serve it's purpose (or maybe I am missing something about approach). For my input image_files:

image_files
['images/image_000.png', 'images/image_001.png', 'images/image_002.png', 'images/image_003.png', 'images/image_004.png', 'images/image_005.png', 'images/image_006.png', 'images/image_007.png', 'images/image_008.png', 'images/image_009.png']
n_views
10

I get:

/mnt/slurm_home/dusan/InstantSplat/utils/sfm_utils.py(65)split_train_test()
-> print(" - train_set_indices: ", train_idx)
(Pdb) n

• train_set_indices: [0, 0, 0, 0, 0, 0, 0, 0, 0, 9]

/mnt/slurm_home/dusan/InstantSplat/utils/sfm_utils.py(66)split_train_test()
-> print(" - test_set_indices: ", test_idx)
(Pdb) n

• test_set_indices: [1 1 2 2 3 4 4 5 6 6 7 8]

[kairunwen]

Hi, @dsvilarkovic.

Split Logic

• Test Set: Selects 12 images, uniformly distributed across the image sequence, excluding the first and last images.

• Training Set: Consists of all images not included in the test set. If n_views is specified, it uniformly samples n_views images from this initial training set.

Example

• For a dataset with 24 images (indexed 0 to 23) and n_views=3:

• Test Set Indices: [1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 22]

• Initial Training Set Indices (Before Sampling): [0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 23]

• Final Training Set Indices (After Sampling, n_views=3): [0, 10, 23]