CLI --affine parameter has no effect

[mstenta]

It seems that the --affine CLI parameter is not working as expected.

Expected behavior

According to the CLI help text for the --affine parameter:

Overwrites multiple parameters to optimize the stitching for scans and images captured by specialized devices.

The tutorial specifically outlines the parameters that should be overridden:

{'estimator': 'affine', 'wave_correct_kind': 'no', 'matcher_type': 'affine', 'adjuster': 'affine', 'warper_type': 'affine', 'compensator': 'no'}

Actual behavior

The CLI script does not explicitly override those parameters when --affine is used, and it explicitly sets them to non-affine defaults.

For example, the --matcher_type CLI parameter sets default=FeatureMatcher.DEFAULT_MATCHER, which is homography.

stitching/stitching/cli/stitch.py

Line 87 in a471af1

default=FeatureMatcher.DEFAULT_MATCHER,

Steps to reproduce

The following Docker commands (with any set of images) will create verbose output with 00_stitcher.txt files that show the actual parameters that are used.

With --affine:

Command: sudo docker run --rm -v ./:/data openstitching/stitch --affine --verbose ./*.jpg

Result: AffineStitcher(**{'medium_megapix': 0.6, 'detector': 'orb', 'nfeatures': 500, 'matcher_type': 'homography', 'range_width': -1, 'try_use_gpu': False, 'match_conf': None, 'confidence_threshold': 1, 'matches_graph_dot_file': None, 'estimator': 'homography', 'adjuster': 'ray', 'refinement_mask': 'xxxxx', 'wave_correct_kind': 'horiz', 'warper_type': 'spherical', 'low_megapix': 0.1, 'crop': True, 'compensator': 'gain_blocks', 'nr_feeds': 1, 'block_size': 32, 'finder': 'dp_color', 'final_megapix': -1, 'blender_type': 'multiband', 'blend_strength': 5, 'timelapse': 'no'})

Without --affine:

Command: sudo docker run --rm -v ./:/data openstitching/stitch --verbose ./*.jpg

Result 00_stitcher.txt:

Stitcher(**{'medium_megapix': 0.6, 'detector': 'orb', 'nfeatures': 500, 'matcher_type': 'homography', 'range_width': -1, 'try_use_gpu': False, 'match_conf': None, 'confidence_threshold': 1, 'matches_graph_dot_file': None, 'estimator': 'homography', 'adjuster': 'ray', 'refinement_mask': 'xxxxx', 'wave_correct_kind': 'horiz', 'warper_type': 'spherical', 'low_megapix': 0.1, 'crop': True, 'compensator': 'gain_blocks', 'nr_feeds': 1, 'block_size': 32, 'finder': 'dp_color', 'final_megapix': -1, 'blender_type': 'multiband', 'blend_strength': 5, 'timelapse': 'no'})

Cause

This is happening because the --affine parameter only affects which class is used (AffineStitcher() instead of Stitcher()), but it does not explicitly override the parameters that are passed into initialize_stitcher().

stitching/stitching/cli/stitch.py

Line 318 in a471af1

stitcher = AffineStitcher(**args_dict)

Solution

Perhaps before line 318 of cli/stitch.py all of the parameters that should be overridden for affine-mode can be reset to the appropriate values.

This would mean that if --affine is used, then the other parameters would have no effect (--estimator, --wave_correct_kind, --matcher_type, --adjuster, --warper_type, --compensator).

The other option is to still allow individual options to be overridden AFTER the --affine overrides are applied, but I feel like someone wants to do that they shouldn't use --affine.

Workaround

For now, if you need the effects of --affine, those parameters can be explicitly set on the command line like so:

--wave_correct_kind no --matcher_type affine --adjuster affine --warper_type affine --compensator no --estimator affine

[mstenta]

Opened a PR that proposes a very simple (one line) fix: #244

[lukasalexanderweber]

wow great catch and awesome issue with really much detail!

That means using --affine mode in the cli was never working :(

I thought that the cli only have those arguments in the args_dict that are actually used by the user.

Anyway, even then if the user passes in --affine --matcher_type homography the matcher_type should automatically be overwritten with affine as you shoud not mix the two models

Certain detailed settings of cv::Stitcher might not make sense. Especially you should not mix classes implementing affine model and classes implementing Homography model, as they work with different transformations.

[lukasalexanderweber]

This is the perfect explanation for #125, since its not affine stitching but homographic stitching with vertical wave correction

[lukasalexanderweber]

closed with da14e49