v0.9.0

MAP-Tk v0.9.0 Release Notes

This is a minor release of MAP-Tk that provides both new functionality and
fixes over the previous v0.8.1 release. However, MAP-Tk is still considered
pre-release software and subject to significant API changes.

Updates since v0.8.1

MAP-Tk Library

• Removed the standalone canonical_transform() function in transform.h in
  favor of an abstract algorithm with implementation in the core plugin.
• Added a function to match_matrix.h to use the match matrix to score each
  track on how important it is in covering the match matrix.
• Moved a function to extract a camera from an essential matrix from the core
  initialize_cameras_landmarks algorithm into epipolar_geometry.h.

Core Plugin

• Added a new algorithm for canonical transform with the implementation
  details moved from the function that was in transform.h.
• Added a new loop closure algorithm implementation that searches for
  matches in a local neighborhood and then establishes keyframes for
  long term loop closure outside that neighborhood. This algorithm
  is generally slower than others, but can be configured to produce
  considerably more dense match matrices with longer tracks. It also handles
  long term loop closure without any assumption about planar scenes.
• Added a new matcher that uses a fundamental matrix constraint to filter bad
  matches.
• Added more config parameters to the initialize_cameras_landmarks algorithm
  to allow more control on how the algorithm is run and what thresholds are
  used.
• Added a new loop closure algorithm that exhaustively matches all pairs of
  frames, or can be configured to match all frames within a moving temporal
  window.

OpenCV Plugin

• Added support for OpenCV version >= 3.0 while maintaining support for
  OpenCV >= 2.4.6. We may drop support for OpenCV 2.X in a future release.
• Overhauled OpenCV algorithm introspection structure. Versions 3.x took away
  the previous ability to introspect algorithm types and parameters,
  requiring, for at least versions 3.x, for there to be manual wrappings of
  each algorithm to be exposed. Removed introspection for any version and
  now have cross-version compatible manual wrappings of all algorithms at
  least present in 3.x that fill the feature detection, descriptor extraction
  and feature matching roles.
• Added an OpenCV fundamental matrix estimation wrapper. OpenCV can estimate a
  fundamental matrix with the presence of outliers using RANSAC.

VXL Plugin

• Added a new algorithm for canonical transform which uses robust estimation
  (e.g. RANSAC, LMS, IRLS) from the Rensselaer robust estimation library
  (rrel) in VXL to fit a ground plane to the data. The ground plane becomes
  the Z=0 plane. The rest of the free parameters are then fit by PCA as in
  the core plugin implementation.
• Added a video_input algorithm implementation that uses the FFmpeg via the
  vidl_ffmpeg_istream class to read video frames and KLV metadata from encoded
  video files.

Visualization Application

• Updated the "Align" tool to use the configurable abstract algorithm
  for canonical transform.
• Added an option to display the world axes around the visible geometry.
  This option can be enabled and disabled in the "View" menu.
• Added an option to export the current scene to webGL. This option is only
  available if VTK was built with the option Module_vtkWebGLExporter turned on.
  This option can then be found in the "File > Export" menu.
• The absolute path to the current frame is now displayed at the bottom of the
  Camera view.
• The GUI now accepts a list of files to be opened on the command line.
• Added a depth map view and a depth map visualization in the world view. The
  depth maps can be thresholded directly in the world view, and the depth map
  view offers several visualization modes such as depths, uniqueness ratios or
  best cost values. Depth maps can be loaded by initializing the parameter
  "depthmaps_images_file" with the path to the depth maps (*.vti files) list in
  the configuration file.

Tools

• Updated the bundle_adjust_tracks tool to use the configurable abstract
  algorithm for canonical transform.
• Added an option to the bundle_adjust_tracks tool to remove all the KRTD
  files from the output directory before writing new ones. This is useful
  when only a subset of cameras are computed and we do not want a mix of
  old and new camera files.
• Added -h option to C++ track features too. Equivalent to --help.

Scripts

• Added a Python script which uses OpenCV to produce mosaicked images of
  video frames using the homographies produced by maptk_track_features.

Default Configuration

• Switched default loop closure method to keyframe based instead of a mix
  of homography based and bad frame detection.
• Switched default feature descriptors to use non-upright SURF.

Fixes since v0.8.1

Build System

• When Fletch is found, or referenced by Vital, CMake will now search the
  Fletch install root when finding packages. This removes the need to
  manually point CMake at each dependent library provided by Fletch.
• Removed CMake option to build shared vs static libraries. Due to the plugin
  infrastructure used in Vital, we can no longer build static MAP-Tk
  libraries.

Plugins

• Corrected all plugins to use the Vital logging framework instead of
  cout and cerr for status, debug, and error messages.

Tools

• Corrected the track_features and bundle_adjust_tracks tools to use the
  Vital logging framework instead of cout and cerr for status, debug,
  and error messages.
• Added a new track filtering option to the bundle_adjust_tracks tool
  that filters tracks based on match matrix importance. The idea is that
  long tracks are usually most important, but some short tracks may be more
  important if they are the only tracks that tie together a pair of frames.
• Removed call to landmark triangulation algorithm that was effectively doing
  nothing.

VXL

• Fixed bug in bundle adjust implementation where input landmarks were being
  accidentally modified in-place.