Change Log

Changed timing routines so you can name each timing block.

Added another add_overlay() overload.

Simplified the add_overlay() interface for image_window.

fixed #include

Added a tool for timing blocks of code

Minor change to avoid a bug in clang 2.9

A few more minor changes to avoid problems with GCC 4.2.1

Minor changes to avoid bugs in GCC 4.1.2 on Debian Etch

I changed the sequence labeling stuff to allow a user defined
sequence type.  So these changes are basically just renaming some
template arguments and minor changes to some typedefs.

Added a unit test for fine_hog_image.

Added fine_hog_image object

Cleaned up the interface for normalization.

Changed poly_image to allow the user to decide if intensity normalization
should be used.

Clarified spec

Clarified spec

Just cleaned up the code a little to make it easier to follow.

Added missing default argument value

Made the object_detector's interface a little more flexible.

Added some member functions to allow the user to set the minimum size
of an image pyramid layer.

Fixed an array indexing bug.

Added some code to help detect and debug improperly coded feature extractors.

Made this code a little more robust.

Improved how matching rectangles are found.  Now get_best_matching_rect()
will use the same matching measure as the rest of the code.

Added a verbose version of setup_grid_detection_templates()

setup_hashed_features() wasn't copying the configuration of the
feature extractor into the scanner.  Fixed this problem.

Clarified spec

Fixed grammar

clarified spec

Added a static const long copy of downsample so you can obtain its value
given only a poly_image type.

Added more unit tests

fixed linker error

minor cleanup

Added the poly_image local feature extractor.

Added the nearest_neighbor_feature_image object.

Fixed formatting

clarified spec

updated a comment

Moved sum_filter() into the file with all the other filtering routines.

Clarified spec

Made the spatial filtering routines return a rectangle which tells you what
part of the image wasn't just zero padding along the border.

Added some overloads of shrink_rect() and grow_rect() which
allow you to adjust the width and height independently.

Renamed a test function so it doesn't show up in the ctags list.

Added the spatially_filter_image_separable_down() routine which is just a version
of spatially_filter_image_separable() that does downsampling at the same time
as it does the filtering.

Made the border_enumerator a little more flexible.  Now the border
doesn't need to be the same size on all sides of a rectangle.

Fixed bugs in this test

Made this code a little more flexible with regard to input matrix type.

Gave the spatially_filter_image routines the ability to add their
results to an output image rather than overwrite the contents
of the output image.

Added an unload() to the hog_image.

Fixed improper normalization in the gaussian() functions.  The
normalization constant was being computed incorrectly.

Reworked the scan_image() routine to make it run faster.  Also added
a new function called sum_filter().

Changed code a bit to make it run faster.

Removed unnecessary load() calls now that they aren't required.

Removed the requirement that an image be loaded before
calling get_best_matching_rect().

Made this object cache image scanner objects internally rather than
reconstruct them for each separation oracle call.

Made this code a little more efficient.

Removed unnecessary requirement on get_block_rect()

Removed cruft

changed code to avoid a compiler warning

Added more serialization tests for dlib::rand.

Changed to avoid compiler error in visual studio 2005

Fixed linker error.

Added the setup_grid_detection_templates() routine.

Added determine_object_boxes()

Made the spec a little clearer.

Made the randomly_sample_image_features() routine a little more general.
Also improved its spec.

Added some more tests and also fixed the existing ones to work
given the recent changes to the object detection tools.

Added the setup_hashed_features() helper routine for configuring
a scan_image_pyramid when it is used with a hashed_feature_image
and projection_function.

Added some useful typedefs to this object.

Added the randomly_sample_image_features() routine.

Fixed constness bug

Clarified spec

Removed constness from the operator() member functions of the object_detector
since they logically modify the state of the object detector now that
we are exposing the state of the underlying image scanner.

Refactored code into a cleaner form.

Made the object_detector's interface a little more complete and flexible.

Made this trainer automatically determine decent parameters
for the test_box_overlap object when it is used and the user
hasn't specifically assigned an overlap tester.

Fixed grammar

The word "overlap" was used to describe the concept of a rectangle
matching with a truth rectangle as well as overlap in the context of
non-max suppression of results.  Since these are two very separate
ideas I renamed some things to avoid any confusion.

Refactored this object so it uses a hashing function supplied
by a template argument rather than its own method.  By default,
this object will use the projection_hash.

Added asserts and filled out the spec files.

Moved create_random_projection_hash() into its own file.

Added some basic locality sensitive hashing tools.

Refined the scan_image_pyramid interface a little.  In particular, I split the
get_feature_vector() method into two separate functions so the interface
is a little simpler and more flexible.

Renamed a variable to make the code clearer.

Improved an error message.

Made the test_box_overlap a little more flexible.  This change breaks
backwards compatibility with the previous version though.

Added an operator != to the file and directory objects.

Added another member function which enables you to get the detection
strengths from the object_detector.

updated spec

Added some simple normality tests for the new rand::get_random_gaussian() method.

Added a function for generating Gaussian random numbers.

Fixed a typo in the spec.

Improved error messages from object detector.

The object_detector had code in it which limited the number of outputs
to 100 rectangles.  This has been removed.

Made assign_image() and assign_image_scaled() capable of assigning to
matrices.

changed code to avoid a compiler warning

Fixed a bug which triggered when using variable length matrices
as the feature vector type.

Changed code so it compiles in visual studio 2010

Changed the add_probability() method of joint_probability_table so
it does a saturating add rather than a normal add.  This ensures the
probability value stays exactly <= 1.  Previously, floating point
rounding error could cause it to be slightly above 1 and would therefore
cause some asserts to misfire during debugging mode.

fixed some typos in assert messages

Fixed a bug which caused the tests to fail if run more than once.

Changed code to avoid a compiler warning

Made the sequence trainer use the reject_labeling() information from
the feature extractor.  Also added the necessary input validation
to make sure this feature doesn't get misused.

Added the contains_invalid_labeling() routine.

Added unit tests for the assignment learning stuff

Relaxed the requirements on the feature extractor interface and also
added some tests to make sure the code really does work with the
relaxed interface.

Added methods to get the assignment by reference rather than returning
a copy by value.

Clarified specs

Clarified spec

Made the order of constructor arguments in the various overloads
for the assignment_function and sequence_labeler consistent.

Filled out the spec and added the usual boilerplate for a trainer object.

removed cruft

Fixed a minor bug in how the cross validation accuracy was being
computed.

Fixed spelling error in comment.

Added a missing assert.

Added a spec for the structural_svm_assignment_problem and added missing asserts.

Added a spec for the assignment problem validation functions and added
missing asserts.

fleshed out this spec

clarified spec

renamed some things

Made the feature_extractor interface a little more efficient.

Setup serialization support for assignment_function objects.

Added initial version of the assignment problem learning code.

Just added some comments about setting up proper input validation
when the user uses the sequence label rejection feature.

Added the is_assignment_problem() and is_forced_assignment_problem() routines.

Clarified spec

Added a test for 64 bit integers.