cbalint13 / opencv-dlco

OpenCV DLCO (Descriptor Learning Using Convex Optimisation)

This is OpenCV implementation of DLCO [1], [2] and [3]. It has booth trainer and descriptor part, reimplemented from original matlab code. Trainer part implemented here uses GPU as mandatory requirement. Descriptor part called VGG was submitted to OpenCV: opencv/opencv_contrib#486

• The minimum hardware requirements for training process are at least 16Gb memory and CUDA capable GPU card with minimum 4Gb of memory. Matrices involved during training reach up to 13Gb, and some validation matrices uses up to 2.5G memory on GPU card for optimised inplace computation. Progress bars indicates realtime computations and estimated time at any point or part of the training process. Core storage containers for the very large matrices uses OpenCV's HDF5 backend, thus these numerical results can be consulted outside the DLCO framework.

• Math part in many places reorders computation from original formulas for being numericaly stable due to large floating point computations & error accumulations. Very special attention was given for these aspects during coding DLCO. Results may differ accross machines and compiler flags. Floating Point HW excepetions are activated to track down potential issues.

• Computations that cannot fit within GPU memory are computed on local CPU using OpenBLAS and OpenMP parallelism. OpenBLAS is also used for its very fast eigen decompositions (SSYREVR) and matrix multiplications (SGEMM). It is recommanded to have OpenMP compilation variant of OpenBLAS for highest possible performance.

• Follow scripts and documentation from the workspace folder for getting through whole training procedure.

Bug reports, suggestions & improvements are welcome !

• Original project: http://www.robots.ox.ac.uk/~vgg/software/learn_desc/

[1] K. Simonyan, A. Vedaldi, A. Zisserman Learning Local Feature Descriptors Using Convex Optimisation IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014

[2] K. Simonyan, A. Vedaldi, A. Zisserman Descriptor Learning Using Convex Optimisation European Conference on Computer Vision, 2012

[3] M. Brown, G. Hua, S. Winder Discriminative Learning of Local Image Descriptors IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011

15-Mar-2016

• Projection matrix learnt over exhaustive 9x45 combinations of mu & gamma regularizers.
• Search space was enough to cover and find best minimizers for 120, 80, 64, 48 dimensions.

03-Mar-2016

• Projection matrix learnt over 9x16 combinations of mu & gamma regularizers.
• Export tool that convert and compact learnt matrices into OpenCV's VGG *.i C code.

18-Jan-2016

• Fixed conv-imgpatches.cpp, very last ending image patch was missing.
• Pool-Region filters re-learnt over 3x256 combinations of mu & gamma regularizers.
• Some older filters are reused from 07-Jan-2016 batch, that process differ by numeric.
• Best selected filters average at 10.682% FPR95 value, but much lower dimensionality.

07-Jan-2016

• Pool-Region filters are learnt over 3x100 combinations of mu & gamma regularizers.
• Best selected filters average at 10.664% FPR95 value, lower than authors exposed.

TODO:

• Hamming distance binarisation part of the DLCO, add it to OpenCV VGG descriptor too. 1