Add image tiling option to ColmapDataParser to use less memory

[couperle]

High resolution imagery can cause OOM during training. The colmap dataparser exposes a downscaling option to mitigate this, but the resulting lack of visual fidelity in the training data may not be desirable. I propose an alternative option: tile the images.

For example, with a tiling factor of 2, each image is split (down the middle in each dimension) into 4 tiles of roughly equal size:

The resulting training requires less GPU memory and runs faster over the same number of iterations. On the nerfstudio poster dataset, there is a noticeable improvement in the quality of the asset:

Default training:

ns-train splatfacto --pipeline.datamanager.cache_images_type=uint8 --viewer.quit-on-train-completion=True --data=data/nerfstudio/poster --output_dir=train --experiment_name=default_pre --max-num-iterations=30000 --timestamp=20200404_000000 --logging.local-writer.max-log-size=0 colmap --orientation_method=none --auto_scale_poses=False --center_method=none --assume_colmap_world_coordinate_convention=False --colmap_path=colmap/sparse/0 --max_2D_matches_per_3D_point=0 --eval-mode=fraction --train-split-fraction=1

Tiling training:

ns-train splatfacto --pipeline.datamanager.cache_images_type=uint8 --viewer.quit-on-train-completion=True --data=data/nerfstudio/poster --output_dir=train --experiment_name=default_pre --max-num-iterations=30000 --timestamp=20200404_000000 --logging.local-writer.max-log-size=0 colmap --orientation_method=none --auto_scale_poses=False --center_method=none --assume_colmap_world_coordinate_convention=False --colmap_path=colmap/sparse/0 --max_2D_matches_per_3D_point=0 --eval-mode=fraction --train-split-fraction=1 --downscale_factor=1 --tiling_factor=2

Exporting assets:

ns-export gaussian-splat --load-config train/default_pre/splatfacto/20200404_000000/config.yml --output-dir output

I also rendered the asset trained with tiling using the original non-tiled cameras to check for artifacts at the seams; there are none.

The default behavior of the colmap dataparser does not change.

[f-dy]

I reviewed this PR. Improvement is clearly visible and measurable.

[brentyi]

It looks good to me as well. Thank you @couperle @f-dy!

[seppestaes]

The comparison on the nerfstudio poster dataset, is it "downscale vs tiling", or rather "no downscale vs tiling"

[couperle]

The comparison on the nerfstudio poster dataset, is it "downscale vs tiling", or rather "no downscale vs tiling"

downscale vs tiling