Impact of Image size on reconstruction

[LaFeuilleMorte]

Hi, I'm trying to reconstruct on an aerial film datasets which has 999 images. My raw image was 4K resolution. I downsized the 4K -> 2K images, and do reconstruction in both 2K and 4K respectively. But the results seemed to be far different:

1. Reconstruction in 4K : 787 images left; Point cloud nums: about 210, 000 .
2. Reconstruction in 2K: 993 images left; Point cloud nuts: about 360, 000

What make so much difference between two resolutions. Please let me know if you have clues on this issue

[jaco001]

After down sample matched points are 'bigger' and less frequent, so Colmap have easier work to match them.

[sarlinpe]

By default, COLMAP resizes images to 3200px max:
https://github.com/colmap/colmap/blob/d065cea09317b2c1cbb9fad66376464119a3ee07/src/colmap/feature/sift.h#L49C7-L49C21
The pixel thresholds throughout the pipeline are thus tuned for this resolution. With very large images, the thresholds would need to be increased. For example:

colmap/src/colmap/sfm/incremental_mapper.h

Lines 69 to 71 in d065cea

	// Maximum error in pixels for two-view geometry estimation for initial
	// image pair.
	double init_max_error = 4.0;

colmap/src/colmap/sfm/incremental_mapper.h

Lines 82 to 83 in d065cea

	// Maximum reprojection error in absolute pose estimation.
	double abs_pose_max_error = 12.0;

colmap/src/colmap/sfm/incremental_mapper.h

Lines 109 to 110 in d065cea

	// Maximum reprojection error in pixels for observations.
	double filter_max_reproj_error = 4.0;

colmap/src/colmap/sfm/incremental_triangulator.h

Lines 54 to 58 in d065cea

	// Maximum reprojection error in pixels to merge triangulations.
	double merge_max_reproj_error = 4.0;
	// Maximum reprojection error to complete an existing triangulation.
	double complete_max_reproj_error = 4.0;

We should have a way to have resolution-agnostic thresholds but this is not yet implemented.

[LaFeuilleMorte]

By default, COLMAP resizes images to 3200px max: https://github.com/colmap/colmap/blob/d065cea09317b2c1cbb9fad66376464119a3ee07/src/colmap/feature/sift.h#L49C7-L49C21 The pixel thresholds throughout the pipeline are thus tuned for this resolution. With very large images, the thresholds would need to be increased. For example:

colmap/src/colmap/sfm/incremental_mapper.h

Lines 69 to 71 in d065cea

	// Maximum error in pixels for two-view geometry estimation for initial
	// image pair.
	double init_max_error = 4.0;

colmap/src/colmap/sfm/incremental_mapper.h

Lines 82 to 83 in d065cea

	// Maximum reprojection error in absolute pose estimation.
	double abs_pose_max_error = 12.0;

colmap/src/colmap/sfm/incremental_mapper.h

Lines 109 to 110 in d065cea

	// Maximum reprojection error in pixels for observations.
	double filter_max_reproj_error = 4.0;

colmap/src/colmap/sfm/incremental_triangulator.h

Lines 54 to 58 in d065cea

	// Maximum reprojection error in pixels to merge triangulations.
	double merge_max_reproj_error = 4.0;
	// Maximum reprojection error to complete an existing triangulation.
	double complete_max_reproj_error = 4.0;

We should have a way to have resolution-agnostic thresholds but this is not yet implemented.

After changing these two configs:
max_image_size = 8192
max_num_features = 15000

I got much more key points and denser point clouds than the original setting. Do I still need to change the pixel threshold in "incremental_mapper.h" and "incremental_triangulator.h" ? I'm not sure if it's significant pixel difference between 3200 px and 8192 px images in clomp.

[ahojnnes]

Closing: resolution-dependent thresholds explained by maintainer.