Change Log

Added draw_surf_points().

Set a default value for the max number of points argument for get_surf_points().

removed unused variable

Made the detection threshold an argument to get_surf_points() rather than having it hard coded. I also increased the default threshold to a more reasonable value. The previous value hardly excluded any points and gave a slightly worse average precision on a standard test dataset.

Made get_surf_points() set the integral image pixel type automatically based on the input pixel type rather than always using long.

Added scope qualifiers to dlib::array so there isn't a name conflict on C++11 compilers.

Added an overload of cca() that can take random_subset_selector objects.

Added to_std_vector() to the random_subset_selector.

clarified cca spec

Clarified spec

To work around a bug in visual studio 2012, I changed the iosockstream from using iostream::tie() to implement its "auto flush on read" feature to use the sockstreambuf's own auto flushing mode instead.

Added an option for a sockstreambuf to automatically flush its output buffers before performing any network reads.

fixed grammar in comment

Added more parallel_for() unit tests and changed code to suppress compiler warnings.

Changed the file and directory objects to not be reference counted. This is so they are safer to use in threaded programs.

Added unit tests for segment_image()

Made segment_image() use the fast version for uint16 pixels.

Generalized segment_image() so it works on any pixel type or array of vectors. I also changed it's interface slightly. In particular, I removed the min_diff parameter and replaced it with an explicit min_size parameter.

Fixed some compiler errors when using certain versions of the verbose parallel_for routines.

Added parallel_for_blocked_verbose() and also slightly improved verbose time remaining estimation.

Fixed a bug in parallel_for() and added unit tests for it.

Added parallel_for_verbose()

Changed default parallel_for chunk size to 8 instead of 4.

Switched randomly_color_image() to use the non-pointer based version of murmur_hash3() to avoid violation of the strict aliasing rule. In particular, the previous version didn't work correctly in gcc 4.7.2 when optimizations were enabled.

Added a non-uniform weighting to each SURF bin. This improves the mean average precision when matching SURF descriptors by about 2%.

Changed the hessian_pyramid so that it has a slightly smaller border region and therefore finds more interest points.

Changed how the surf descriptor is computed slightly to improve its accuracy.

Made average_precision() a little more generalized.

Changed ranking evaluation functions to return the mean average precision in addition to just raw ranking accuracy. This changes their return types from double to matrix<double,1,2>.

Added tests for average_precision()

Added average_precision()

All I did was put an "if (NOT TARGET dlib)" around the main body of the CMakeLists.txt so that we avoid a common cmake warning.

simplified cmake file

Simplified cmake scripts

Added a simple cmake script users can use to simplify their own CMakeLists.txt scripts

Added unit tests for new matrix operator>> and fixed a bug in it.

Added an operator>> for matrix objects which allows you to read in ASCII matrices using the format used by operator<<.

clarified spec

Fixed a bug in the image_window and image_display GUI tools which caused them to not redraw overlay lines correctly in certain cases involving non-default zoom levels.

A minor change to avoid a compiler warning in visual studio.

Added a pipe test that catches the bug I just fixed.

Fixed a bug in the Windows implementation of the signaler object, which was found by Isaac Peterson. The bug caused the program to deadlock if signal() or broadcast() was called at exactly the same time a wait_or_timeout() funciton timed out.

There was a bug in the version of draw_line() that draws directly onto an array2d type image (not the one that draws onto a GUI canvas object). The bug triggered whenever a perfectly horizontal or vertical line that extended outside the image was drawn. This has been fixed.

Clarified spec with regards to infinite weight edges.

Added negative_dot_product_distance and also clarified spec regarding thread safety.

Added cosine_distance.

Added missing asserts

Simplified the code a little by replacing some stuff with calls to parallel_for_blocked()

Added more notes to spec

Added parallel_for() and parallel_for_blocked().

made test more robust

Made the distributed structural svm tools use the same improved job/buffering rules as I recently added to the structural_svm_problem_threaded object.

Made svm_struct_controller_node support network_address objects.

Added comparison operators for network_address.

Fixed bug in the axpy/scpy BLAS bindings which triggered when empty matrices were assigned to each other.

Fixed a bug introduced yesterday

Made tests more robust

Changed code to avoid unnecessary cutting plane copying. This makes the optimizer noticeably faster in some cases.

Added missing const keywords

Modified structural_svm_problem_threaded to reduce the amount of copying overhead.

Added BLAS bindings for xscal and xaxpy routines.

Added another folder to blas search paths

Changed the code so that each worker thread gets more things to do for each job. This should give better overall CPU utilization than the previous smaller job sizes.

Expanded the set of folder we search when looking for BLAS and LAPACK.

clarified spec

Added count_bits() and hamming_distance()

Changed general_hash so it uses dlib::hash to hash integers rather than returning them unmodified.

made unit test more robust

clarified spec

Added more cca() unit tests.

Made svd_fast() accept a wider range of matrices as arguments.

Made some of this code a little more general.

Increased the default number of power iterations done by svd_fast() from 0 to 1.

Made the svd_fast() code a little more readable and memory efficient. Also added the orthogonalize() function.

Added a version of get_q() to qr_decomposition that allows the user to get the Q matrix by reference rather than by value.

Added make_sparse_vector_inplace()

refined spec a little

renamed function

Refactored the Platt scaling code a little. Now there is a function, learn_platt_scaling(), that allows you to directly call the Platt scaling code without supplying a trainer object.

fixed a typo

Switch a DLIB_CASSERT to DLIB_ASSERT

Added an overload of get_next_double_click() that allows the user to find out which mouse button was double clicked.

Fixed a problem where the pixel click location reported by image_window::get_next_double_click() was slightly off.

Changed the image_window so it doesn't try to adjust the size or position when set_image() is called with an image the same size as the one currently being displayed.

Fixed a bug which could potentially occur when empty std::vector<char> or std::vector<unsigned char> were serialized.

Added serialization support for std::vector<bool>.

Minor change to avoid a compiler error in 4.1

Added some missing includes that only triggered a compiler error in clang and also fixed a function call ambiguity.

made tests more robust

Added unit tests for cca() and svd_fast()

Made cca() more numerically robust.

Added missing asserts

Added cca()

Added svd_fast()

clarified spec

Added some features to the image display widgets to let the user easily get information about where the user is clicking.

Fixed spelling error in comment

Fixed an endianness bug in the PNG I/O functions which occurred when 16bit grayscale PNGs were used. libpng doesn't automatically convert from host endianness to big endian as the PNG standard demands. You have to explicitly tell it to do this or it will write out the 16bit pixel data in host order which result in the wrong thing happening on little endian chips. Similarly for reading PNGs.