sift_anatomy_20140420_extra.tar.gz
 - first to contain describe_keys.c

sift_anatomy_20140530.tar.gz
 - Review pascal Getreuer part 1.
 - const variable
 - lib_util.c (max tab)..
 - assert


sift_anatomy_20140609.tar.gz
 - Review Pascal Getreuer part 2.

// 20140609
- Bug in the normalization of the feature vector.
This is not bin*255/max(bins) but bin*512/norml2(bins)

// 20140611
- sqrt(2) replace by a define
- bug in assert tests comparing floats that can be equal
  correction use FLOAT_EPSILON to include the equality case.
  in lib_descriptor in assert M_PI
- M_PI was defined twice with various number of decimals
- fmod to express the gradient in 0 2pi (atan2 gives in -pi pi)
   the problem fmod(x+10*M_PI,2*M_PI) -> fmod(x+2*M_PI,2*M_PI)
- Extrema interpolation. Possible difference between anatomy and lowe binaries.
  If the offset in scale is larger than 0.5 (the half distance between two
  scale samples) Lowe doesn't (I believe) jump to the next scale. Only, change
  sin position are implemented.  Observations: With this sampling parameters
  the threshold over the offset has an impact on the keypoint location (number
  of keypoints position) Visually, our results are more similar to Lowe's if we
  don't

// 20140616
 - in a few assert. replaced FLT_FLOAT with a larger constant EPSILON.
 - in lib_util linear_conversion : response in [0,1]. skip the normalization if the vector
   is constant (0 division)
 - array_max
 - array_min
 - find_array_max
 - fixed index in description  (s+1) -> s. This is the index used to chose the
   layer of the octace (gradient).

// 20140618
 - bug, find two first minima of an array
 - new functions in lib_util.c (euclidean_distance  euclidean_distance_square)
 - TODO, the comparison should be performed on the squared euclidean distance
   (for speed and precision).  tested it gives the same result than euclidean
   distance and exactly the same result as Lowe's code (siftDemoV4/match.c).

// 20140630
 - in lib_sift.c replace hardcoded default parameters with a call to the
   sift_assign_default_parameters function
 - in demo_extract_patch.c  remove printImage definition
 - in demo_extract_patch.c  function thumbnail_ori rewritten
 - in demo_extract_patch.c  remove not used functions gradient_field_ori() and
   gradient_field_descr()
 - sprintf() is replaced with snprintf() The functions snprintf() and
   vsnprintf() write at most size bytes (including the terminating null byte
   ('\0')) to str.
 - remove 'count' output for functions accumulating histograms,
   sift_extract_feature_vector and sift_accumulate_orientation_histogram,
   they now return void

// 20140801 - Major change in the interface
 sift_cli
 - new parameter parser pick_option() (inspired by enric's implementation).
 - each parameter is adjustable individually
 - the help menu and README are changed accordingly
 TODO:
 - add an option to do only detection
 - remove a few useless output files (extra_????.txt), namely ..REJ.. ..MULT..
 - euclidean_distance() uses euclidean_distance()
 - replace ceil by floor+1 !!! in lib_description.

// 20140808 _Leo_F_
 - new function fprintf_one_keypoint(). This will guaranty consistency between sift_cli and match_cli.
 - refactoring match_cli, the useless 'pair' structure is replaced by three sift_keypoints structure and a list of doubles
 - sift_keypoints structure: the list of pointers to keypoint structures is replaced by a list of keypoints.

// 20140819
 - fixed BUG segmentation fault (realloc_sift_keypoints - bad comparison size and capacity)
 - fixed memory leaks in filtering question (the copy of the keypoint was allocated regardeless of the test results)
 - fixed BUG affecting sift_find_ori_and_fill_descriptors and sift_anatomy_orientation_and_description:
         sx instead of sy in keypoints_attribute_one_orientation

// 20140901
 - two example codes that use the lib_sift.o: example.c and sift_cli_standard.c
   (requested by the reviewer)
 - unnecessary casts from (void*), as in the return value of malloc, can be
   avoided.
 - UNDONE printing the keypoints starting with keypoints with the largest
   scale.
 - FIXED BUG in the orientation histogram peak interpolation. The wrong formula
   led to a quantified set of principal orientations.

// 20140909
 - computation of number of octave (now put aside in a static function)
 - threshold adaptation to scalespace discretization (now put aside in a static function)


// 20140911
 - update io_png.c (it includes the string.h library that defines strcmp).
 - anatomy2lowe script to convert orientations and descriptors
 - FIXED BUG in match_cli, a segfault was raised in the absence of any match.
 - Check if all input filenames are valid


// 20140929
 - TODO in lib_keypoint.h definition of the 'val' field/ val is not normalized, remove wrong comment.

// 20141002
 - Add 'iter' in struct keypoint (number of iterations).
 - When the discrete extrema is on the edge of the image, the process is stopped.
 - TEMP, no jump in scale
 - DEBUG, nmber of iteration was zero sometimes (impoossible) 
 - The definition of the DoG operator is function of the sampling parameter ss_nspo.
   A new parameter (dog_nspo) allows to define  the DoG operator without changing the discretization (gradual)
 - In gradual, all scales in the DoG scalespace are computed from the scalespace seed ( sigma_min).

// 20141007
 - Three floating point precision (single double quad, 32 64 128) defined in
   lib_util.h preprocessor definition QUAD DOUBLE
 - Extract a fiber of a scalespace on a set of point



// 20141104
 - DCT everywhere
 - gradual computation of the scale-space for low memory usage


// 20141111
 - extra output: 3D Hessian and gradient + the 27 neighbors for each keypoint.
   Used to assess how well the interpolation system is conditionned.


// 20141201
 - FIXED BUG in sift_cli (float to _myfloat conversion missing)
 - ADDED to see 33x33 patches surrounding a given set of extrema in the
   DOG scalespace.
   extra_sift_cli_extrema_patches
 - ADDED (TEMP ?) - discard interpolation if ofset >0.5  (instead of 0.6).
   This will guarantee that (o,s) extrapolated from the keypoint sigma is the
   (o,s) coordinate of the first discrete extrema.

 - FIXED  the read_keypoint + conversion to  anatomy_keypoints takes default parameter values.
 - flag_dct reverse.

// 20141216 - UY
 - FIXED BUG: flag_dct inverted in extra_sift_cli
 - extra_sift_cli stdout with the 27 neighbors
 - with itermax = 0, the binaries now extract the discrete extrema.

 - TODO   FIX BUG with ellipses/ normalized_ellipses/
 - TODO   FIND routine to print affine normalized patches.

// 20141217
 - EPSILON was set to 0 - It's now set to FLT_EPSILON (for doubles).
   less detection artifact on  synthetic images.
 - TODO / replace with a justified value of epsilon (this one is too short
        or replace EPSILON with a threshold parameter.

...

// 20150415
  - flag to compute an extra auxiliary image below ssmin (warning: ssin = smin
    and is a forbidden configuration and will produce a failure).


// 20150517
 - addoutput the offset

// 20150531
  - fixed the scale and octave output for gradual implementation