[ Up ]

Alignment

Alignment compares each supplemental frame in the sequence with an intermediate incremental rendering known as the alignment reference image. Based on this comparison, a transformation is assigned to the supplemental frame. Many different transformations can be assigned before alignment of the frame is complete.

This manual page offers an overview of alignment options; linked from this page is a sub-page offering more detail on transformation data files.

Transformations

The variables used to adjust alignment are called transformations; they map points from the source image coordinate system to a target coordinate system (the rendering coordinate system).

ALE supports three transformation classes -- translational, euclidean, and projective -- as well as barrel/pincushion distortion correction. The translational class applies only translations to the inputs, changing only the positions of images; the Euclidean class applies translations and rotations; and the projective class applies general projective transformations. Additionally, barrel/pincushion distortion correction can be applied to each frame. The Euclidean class is most appropriate for use with scanners, or as a first pass for projective transformations. The projective class and barrel/pincushion distortion correction are most appropriate for use with cameras. Except when capturing flat scenes, ALE does not correct perspective changes, so movement of cameras should ideally be constrained so that no large translations occur.

Transformation Class Options

--translation     Only adjust the position of images
--euclidean       Adjust the position and orientation of images [default]
--projective      Use projective transformations.  Best quality, but slow.

Transformation data file options

Transformations used in alignment can be loaded from, or saved to, a transformation data file. This can be useful when performing alignment in several passes, or when refining rendering options. A transformation data file is required for barrel/pincushion distortion correction.

--trans-load=x    Load initial transformation settings from file x
--trans-save=x    Save final transformation data in file x

Alignment following

The --follow option hints that frames tend to be more closely aligned with adjacent frames in the sequence than with the original frame. This hint can be useful even when initial alignment information is loaded from a file.

--identity        Frames align closely with the original frame.  [default]
--follow          Frames align closely with their immediate predecessor.

Match statistics

The match statistic indicates how well a transformed frame aligns with the alignment reference image; it is used to select or reject transformations during alignment. After alignment of a particular frame is complete, ALE displays the corresponding final match statistic. If the value is close to 100%, then the frames are well aligned; very low values can indicate misalignment; but even frames that are very well aligned typically do not achieve 100% alignment. For improving image quality, ALE works most effectively when match values are lower than 100%.

Match threshold

A match threshold can be specified; no images with final match statistics falling below this threshold will contribute to the final output.

--threshold=x     Min. match threshold; a perfect match is 100.  (0 is default)

Alignment failure

When an image fails to meet the match threshold, a transformation is still assigned to the frame (e.g., for writing to a transformation data file). The assigned transformation can be either the optimal alignment found (this is default) or the default transformation.

--fail-optimal    Frames beneath threshold are aligned optimally.  [default]
--fail-default    Frames beneath threshold keep their default alignment.

Alignment error metric exponent

The function calculated at each pixel to determine the match statistic is called the alignment error metric, and is of the form (a-b)^x. The value x is called the error metric exponent, and is 2 by default. Larger numbers indicate that alignment will be more influenced by smaller image features.

--metric=x        Set the alignment error metric exponent.       (2 is default)

Alignment Channel Options

In calculating the per-pixel error metric, there are three ways in which ALE can handle color channels. By default, ALE adds the channels before calculating the match, but it can also be configured to rely solely on the green color channel, or to use all three channels separately.

--align-all       Align images using all color channels
--align-green     Align images using the green channel
--align-sum       Align images using a sum of channels [default]

Monte Carlo Alignment

Aligning large images can take a very long time if all pixels are examined in determining the match statistic, so it is often desirable to examine a smaller subset of pixels. The Monte Carlo alignment option allows this. The number of pixels used is specified as a percentage, and smaller numbers usually mean faster, but less precise, alignment. For defaults, see the default settings page.

--mc <x>          Align using, on average, x% of available pixels (0 < x < 100)
--no-mc           Align using all pixels.

Alignment weight map

Typically, measured differences at each pixel in the alignment reference image contribute equally to the match statistic. To weight pixels differently, a weight map can be used. The scale of the map should correspond with that of the alignment reference image, and if the upper-left corner of the map does not coincide with the upper-left corner of the first frame in the sequence, then a non-zero offset should be specified to indicate the difference. See also exclusion regions.

--wm <f> <x> <y>  Use weight map image <f> at offset (<x>, <y>)

Frequency weighting

If ALE is compiled with FFTW support, then a high-pass-filtered version of the alignment reference image can be used to weight reference image pixels' contributions to the match statistic. Frequency limiting can occur as a fraction of the highest horizontal frequency, the highest vertical frequency, or the highest average frequency. Limit values should range between 0 (pass all) and 1 (pass none). If any limit is assigned a value of 1, no alignment will occur, since all frequencies will have been excluded. An output file can optionally be specified for high-pass filtered data.

--fl <h> <v> <a>  High-pass filters: horizontal <h>, vertical <v>, average <a>.
--flshow <o>      Write high-pass filtered data to file <o>.

Algorithmic weighting

This option writes the current alignment reference image and definition map to specified files, executes a specified program with these arguments, and then reads back alignment weights. This option requires POSIX libraries, and requires that ALE have been built with POSIX support enabled.

--wmx <e> <r> <d> Write reference <r>, definition <d>, execute `<e> <f> <d>`,
                  read weights <r> back.

Perturbation

The perturbation size determines the magnitude by which alignment parameters are changed. The size is initially set to be large, allowing large changes in alignment to be evaluated, and is reduced as the locally optimal alignment is found for each size. When this size drops below a specified lower bound, the frame is considered to be aligned.

Perturbation bounds

These options determine the upper and lower bounds for perturbation size. The perturb-upper and perturb-lower bounds apply to rotation (in degrees), translation (in pixels), and the movement of the boundaries of a projected frame (in pixels). By appending the '%' character, values may optionally be specified as a percentage of the number of pixels in the smallest image dimension (0.7.1-patch1 and later). The rot-upper bound disables rotational perturbation above a certain perturbation size (in degrees). To disable alignment, set perturb-upper to zero.

--perturb-upper=x Perturbation upper bound in pixels/degrees   (14% is default)   [0.7.1, and earlier, default to 32.0]
                     ('x%' uses a fraction of the smallest image dimension.)      [0.7.1-patch1 and later]
--perturb-lower=x Perturbation lower bound in pixels/degrees  (.125 is default)
                     ('x%' uses a fraction of the smallest image dimension.)      [0.7.1-patch1 and later]
--rot-upper=x     Rotation-specific perturbation upper bound  (32.0 is default)

Barrel/pincushion distortion adjustment multiplier and rate

When using barrel/pincushion distortion parameters, any frame-to-frame adjustments to these parameters are made in a manner dependent on the perturbation size. A multiplier, the barrel distortion adjustment multiplier, is used to determine the perturbation of these parameters, based on the generic perturbation size. Additionally, the frame-to-frame rate of change of each barrel distortion parameter can be limited to a specified maximum. To disable frame-to-frame adjustment of barrel/pincushion distortion parameters, set bda-mult to zero. Setting bda-rate to zero disables rate limitation.

--bda-mult=x      Barrel distortion adjustment multiplier   (0.0001 is default)
--bda-rate=x      Barrel distortion rate of change maximum  (0.0004 is default)

Level of detail

Alignment at large perturbation sizes is usually carried out on reduced-detail images. To disable this, set lod-max to log₂(perturb-upper).

--lod-max=x       LOD scale factor is max(1, (2^floor(x))/perturb)  (1 is def.)

Perturbation Types (experimental; 0.7.1-patch1 and later)

Perturbations are normally performed in output image coordinates. To perform perturbations in source image coordinates, specify --perturb-source.

--perturb-output  Apply perturbations in output image coordinates. [default]  [0.7.1-patch1 and later]
--perturb-source  Apply perturbations in source image coordinates.            [0.7.1-patch1 and later]

Global Searching (experimental; 0.7.0-patch1 and later)

In cases where important image features do not have sufficient overlap using default alignment parameters (determined by --identity and --follow), a number of translations can be applied to the default parameters in order to find a new alignment starting point. Since this process searches the entire region of the alignment reference image, it is called a global (as opposed to local) search. To avoid misalignment, a minimum overlap area can be specified [0.7.1-patch2 and later].

--gs <type>       Set global search to <type>, one of:                             [0.7.0-patch1 and later]
                     local     Local alignment only [default]
                     inner     Alignment reference image inner region
                     outer     Alignment reference image outer region
                     all       Union of inner and outer
                     central   inner if below threshold or better; else, outer.

--gs-mo <x>       Set <x> pixel min. overlap for global search. (16 is default)    [0.7.1-patch2 and later]

Verbatim copying and distribution of this entire article is permitted in any medium, provided this notice is preserved.