Fix/vl feat compliance3

[griwodz]

Description

This PR merges PRs #114, #115 and #128, making it easier for those who need it.

It adds two things:

• faster grid filtering done on the CPU
• results of brute-force GPU matching between two sets of descriptors, using the new method int3* FeaturesDev::matchAndReturn(FeaturesDev*), which are available both on the CPU and GPU side

From PR #114, it fixes a bug in L2 normalization.

From PR #114, it gets a descriptor computation that is compliant with VLFeat. The original PopSift descriptor was not compliant because it used the angle differently: while VLFeat spreads the angular part of the descriptor over up to 8 bins of the descriptor's histogram, PopSift would only spread it over two. Furthermore, descriptors are arranged differently into the histograms (according to the VLFeat documentation, VLFeat does that also differently from Lowe's original code and Changchang Wu's SiftGPU).

From PR #115, it gets a test script that runs VLFeat and PopSift on one or more images and creates the descriptors with orientations. compareSiftFiles.cpp performs a brute force CPU-sided comparison of the output files and outputs sorted files and some distance computations. These are then transformed with bash and awk and rendered visually with gnuplot.

From PR #128, it gets the ability to ignore Edge and Peak Thresholds. These are ignored when their respective configuration variables are set to 0.0f or less.

This PR adds grid filtering that does now rely on CUDA Managed Memory and sorts initial extrema on the CPU side. This is faster than the 2 former approaches that needed a lot of memory management to do it on the GPU side. Grid filtering makes it possible to overlay the image with a grid and ensure that similar numbers of extrema are used from each of the cells, up to an overall maximum. Grid filtering code is excluded if the CMake variabe PopSift_USE_GRID_FILTER is set to OFF.

Features list

• Command line and Config settings to use VLFeat-compliant feature extraction.
• Scripts that compare VLFeat-compatible PopSift behaviour with original VLFeat behaviour. A brute force CPU tool comparing two files containing SIFT descriptors. VLFeat and Matlab syntax supported (132 and 133 numbers per line, respectively).
• Ignoring two checks when the respective Config values are 0 or less.
• Added getDefault functions and usage descriptions to the Config class.

Implementation remarks

The threshold for accepting an orientation in VLFeat is computed before smoothing the histogram, PopSift did it after computing the histogram (consequently accepting fewer orientations). This was changed to work like VLFeat. This change is not restricted to modes that try to be as close to VLFeat as possible, but used for all modes.

For backward compatibility, the VLFeat descriptor is not the default descriptor. The following settings lead to results that seem to be identical to the VLFeat 0.9.20 C code and command line (the 0.9.21 command line does not work):

--desc-mode=vlfeat              : use the VLFeat descriptor extraction
--norm-mode=classic             : use L2 normalization, not RootSift
--norm-multi 9 --write-as-uchar : multiply descriptor values with 2^9 and cast to uchar
--write-with-ori                : output of descriptors in the same format as VLFeat command line (instead of OpenCV-compliant)
--initial-blur 0.5              : PopSift default
--sigma 1.6                     : PopSift default
--threshold 0                   : set the peak threshold to 0
--edge-threshold 10.0           : PopSift default

[tkircher]

Adding #include <thrust/host_vector.h> to s_filtergrid.cu fixed the one and only compilation issue for me, just in case you were wondering if it was actually necessary for anyone.

[griwodz]

@tkircher Thanks! I'm pretty sure that this PR is a big improvement over develop. It may not be absolutely 100% identical to vlfeat yet, but the descriptors including the statistical distribution of their histograms look really similar now. There's also the minor normalization bug fixed and test code.