Skip to content
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time




Engin Tola



web   :
email :



See licence.txt file.



DAISY is a local image descriptor designed for dense wide-baseline matching
purposes. For more details about the descriptor please read the papers:

  author = "E. Tola and V. Lepetit and P. Fua",
  title = {{DAISY: An Efficient Dense Descriptor Applied to Wide Baseline Stereo}},
  journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
  year = 2010,
  month = "May",
  pages = "815--830",
  volume = "32",
  number = "5"


  author = "E. Tola and V.Lepetit and P. Fua",
  title = {{A Fast Local Descriptor for Dense Matching}},
  booktitle = "Proceedings of Computer Vision and Pattern Recognition",
  year = 2008,
  address = "Alaska, USA"



There is a main.cpp file under ./src/ directory as an example code. You can
inspect it on how to use the daisy class. However, this software is intended to
be used as a library and has additional operation modes in that form.

1. To build the example application code run


   (look at EXAMPLES section below)

2. To build the library, remove src/main.cpp from the 'sources' list in the
makefile. Run 'make library' for a static library and 'make slib' for a shared

This will generate the libdaisy.a ( for shared library) file under
./lib directory. To install this to your system, run

   make install-lib  ( or make install-slib for shared library)

This will install the library to 'installdir' in your makefile. This command
also generates the daisy.pc file which is installed to
'$(installdir)/lib/pkgconfig/'. The pkg-config[1] utility can be used to include
daisy to your projects with

   CFLAGS+=`pkg-config --cflags daisy`
   LDFLAGS+=`pkg-config --libs daisy`.

In your source file, include 'daisy/daisy.h' to use the library.

3. The code is documented as per Doxygen[3] standards and a Doxyfile is supplied
in the library. To generate the documentation, run

   doxygen Doxyfile
   make dox

You can reach the html documentation under './doc/html/index.html'

It is possible to use opencv[2] with daisy. Opencv implementation of the
convolution is faster than the one used by default in the library. Apart from
this, the library is free of opencv. Opencv is included/linked using the
pkg-config utility and if you don't have pkg-config you should either install it
or change the makefile accordingly. To enable opencv, uncomment the
'define_flags' and 'external_library' lines in the makefile.

4. To enable debugging for the library change the 'optimize=true' to
'optimize=false' in the makefile.

5. Library can be compiled with OpenMP[4] for parallel processing. To enable
this, change 'parallelize = false' to 'parallelize = true' in the makefile and
enable -DUSE_OPENMP flag for 'define_flags'. When the optimize flag is set to
false then parallelization is disabled also.

[1] pkg-config :
[2] opencv     :
[3] doxygen    :
[4] openmp     :




The library has two operation modes:

MODE 1. In the first case, you can precompute the descriptors at every point with

   daisy* desc = new daisy();
   desc->verbose( verbose_level ); // 0,1,2,3 -> how much output do you want while running
   desc->set_parameters(rad, radq, thq, histq); // default values are 15,3,8,8
   desc->compute_descriptors(); // precompute all the descriptors (NOT NORMALIZED!)
   // the descriptors are not normalized yet

and get the specific descriptor at a point (y,x) with

   float* thor = NULL;

Here, you can get the descriptors at only integer y,x values.

MODE 2. You can also precompute orientation layers with

   daisy* desc = new daisy();
   desc->verbose( verbose_level ); // 0,1,2,3 -> how much output do you want while running
   desc->set_parameters(rad, radq, thq, histq); // we use 15,3,8,8 for wide baseline stereo.

a) get the descriptor at floating point locations with any orientation as

   float* thor = new thor[desc->descriptor_size()];

   desc->get_descriptor(y,x,orientation,thor); // returns normalized descriptor


   desc->get_unnormalized_descriptor(y,x,orientation,thor); // unnormalized

   see NORMALIZATION section for details on normalization.

   orientation in [0 360)

b) or get the descriptor with a warped grid

   float* thor = new thor[desc->descriptor_size()];
   desc->get_descriptor(y,x,orientation, H, thor);

   here H (float H[9]) is a homography matrix used to warp the grid of daisy and
   (y,x) is on the unwarped image. You can use this function in the case where
   you have two images and you want to compare the descriptors of two
   corresponding points which are related with a planar transformation. The
   transformation is encoded with a Homography matrix.

   orientation in [0 360)




As of version 1.5, by default, descriptors are not normalized in MODE 1. You
need to call normalize_descriptors() in order to apply normalization. For MODE
2, you have two options:

(a) get_unnormalized_descriptor
(b) get_descriptor

function (a), as per its name, returns an unnormalized descriptor and function
(b) returns the normalized descriptor.

You can call set_normalization() function with NRM_PARTIAL, NRM_FULL and
NRM_SIFT flags to change the applied normalization algoritm. For example,


is equivalent to


which is equivalent to


set_normalization() function changes the used algorithm globally whereas you can
apply individual normalizations using normalize_descriptor() function.

The default algorith is NRM_PARTIAL. Possibilities are:

NRM_PARTIAL: Each histogram is normalized independently so that their L2 norm is 1.
NRM_FULL   : The whole descriptor is normalized so that its L2 norm is 1.
NRM_SIFT   : The whole descriptor is normalized recursively so that its L2 norm is
1 and no individual value is bigger than m_descriptor_normalization_threshold =
0.154 as in SIFT.




While computing the descriptors, by default, interpolation is used. However, it
is possible to disable the interpolation using the disable_interpolation()
function. This will decrease the computation time at the expense of some minor
performance loss. You can alleviate this by using the extra_sub_layers() and
upscale_image() functions. I recommend to disable the interpolation in only
cases where you run DAISY in MODE 2 and you have to compute the same descriptor
at many different orientations many times ( i.e. in a dense stereo problem, we
test many different depths at many different orientations due to epipole
change. this causes the use of excessive interpolation and it starts to become a
performance issue. )




It is possible to set the memory of the library from outside to let the user
call the library in a loop without allocating/deallocating the necessary
space. The functions for this are :

  void set_descriptor_memory(float *descriptor, long int d_size)
  void set_workspace_memory (float *workspace,  long int w_size)

you can use compute_workspace_memory() and compute_descriptor_memory() functions
to get the required memory; i.e,

  wsz = desc->compute_workspace_memory();
  float* workspace = new float[ wsz ];
  desc->set_workspace_memory( workspace, wsz );

if you're not running the library in MODE_1 you don't need to set the descriptor
memory. In fact, you should not set it cause it'll be a waste.  YOU SHOULD CALL




You can use the executable for

1.a. computing the descriptor of a single point y,x=(45,132) with orientation 35

   ./daisy -i image.png -d 45 132 35

1.b. computing the descriptor of a single point y,x=(45,132) with orientation 35
without using interpolation

   ./daisy -i image.png -di -d 45 132 35

2. computing all of the descriptors and saving them in ascii

   ./daisy -i image.png -sa

3. computing all of the descriptors and saving in binary with NRM_FULL normalization

   ./daisy -i image.png -sb -nt 1

   ps: you can use the load_binary function in kutility/fileio.h to load the
   saved descriptors like
      float* descriptors = 0;
      int h,w,nb;
      load_binary(file, descriptors, h, w, nb);

4. making a time run

   ./daisy -i image.png -tr


Have Fun!
Tuesday, August 18, 2009 13:27:58 +0200


DAISY: A Fast Local Descriptor for Dense Matching




No releases published


No packages published
You can’t perform that action at this time.