[ptx] Release: autopano-sift 1.5: bug fixes and exciting new features

[Sebastian Nowozin]

Hello everybody,

after a long time travelling and being unable to respond to all the mails, I
finally found time to improve upon autopano-sift and fix the bugs Alexandre
reported. Thanks for the reports, it really improves the results. Besides
various other smaller improvements detailed in the changelog below, I put in
three exciting new features, which I hope will improve the results a lot:

 1. Automatic pre-aligning of images
    In case the images fullfil some simple criterias (first images lie in a row
    on the horizon in either strict left-to-right or right-to-left order), all
    the images can be pre-aligned so the panotools optimizer has a good minima
    to start from instead of trying rather randomly. Also includes a "bottom
    estimator", that guesses about where the bottom of an image is based on
    keypoint density. You can override the use of that by using
    --bottom-is-left or --bottom-is-right for 90/-90 degree rotated images. The
    automatic estimation works well for landscape photography.

 2. Automatic horizon generation
    Under the same conditions as in 1., automatic generation of horizon lines
    can be enabled.

 3. Component enumeration
    After matching, all connected components are enumerated and shown. This
    way, you can quickly see which image groups are failing to match and can
    either increase keypoint density by increasing the downscaleResolution, or
    by manually fixing the components together.
    Question: does hugin's wizard check for components or only single images?


The release is available at the usual location:

	http://user.cs.tu-berlin.de/~nowozin/autopano-sift/

Please test and let me know of your results, but do not expect a reply before
September, as I will be travelling again. I promise to check every report I
receive in the meantime though.


The detailed changelog follows:

autopano-sift 1.5, libsift 1.3
2004/08/06

  * Fix two severe mathematical bugs found by Alexandre Jenny
    (autopano at kolor dot com). Thanks! One is a wrong factor in a matrix
    calculation that lead to fewer keypoints being detected. Another is an
    algorithmic bug that also reduced the number of keypoints. This bug has
    previously been found by Willi Nagel (lascobar at gmx dot de), but I did
    not recognized it back then, sorry. Both bugfixes together double the
    number of keypoints generated.
  * Fix documentation of generatekeys, where the downscaleResolution option
    was not explained correctly.
  * Change behaviour of the downscale resolution in the generatekeys utility.
    By supplying zero as downscale resolution, the image is taken verbatim,
    without any scaling.
  * Fix small bug in autopano GUI: images are no longer added when cancel is
    hit in the file selection dialog.
  * Fix a rare bug where a RANSAC model fit was accepted that was not fitted.
    This occured only when at two or more scales the same point location was
    selected to fit a model.
  * Change the Best-Bin-First search cutoff depth from static 130 to
    logarithmically increasing, depending on the overall keypoints (formula
    is (ln(keypoints)/ln(1000))*130. This increases accuracy, but introduces a
    logarithmic factor into the otherwise linear time complexity.
  * Change the autopano-complete.sh script so that a downscale resolution of
    800 pixels is used in case none is given. Any given railable on the command
    line utility (autopano.exe) and has to be enabled through the "--align"
    option. See the manpage for further information on restrictions on images
    that can be used with this option. This computation comes at virtually no
    costs and everybody should try it out :-)
  + Add the "--bottom-is-left" and "--bottom-is-right" options for use with
    "--align" option, in case 90/-90 degree tilted images are not properly
    recognized.
  + Add experimental horizon generation. This works only in conjunction with
    the "--align" option and simply puts up to a given number of horizontal
    control point lines on the images found to be in the first row.


Thanks for testing,
Sebastian

-- 
nowozin at cs.tu-berlin.de --- http://user.cs.tu-berlin.de/~nowozin/