1-bit news

OK, it’s been a while again, but here are the reasons for being silent. 1-bit LBD reconstruction is out! Remember all those posts about LBDs, FREAKs, BRIEFs, and reconstructing images ? The latest version of tis work is out with a major feature: 1-bit quantized LBDs reconstruction. Yes, 1-bit ! To achieve this, we leverage some results from 1-bit Compressed Sensing thanks to my co-author Laurent Jacques. So, the pre-print is on arxiv and on infoscience, and you can get the code from github.

Linearized model of LBPs

LBP model An LBP can be decomposed into two tiers: first, a real description vector, obtained by convolution-then-difference then the quantization (binarization) operation. Mathematically, the real i-th component of the descriptor is computed with the formula: $${\mathcal L}(p)i = \langle{\mathcal G}{x_i, \sigmai} , p \rangle - \langle G{x_i’, \sigmai’} , p\rangle, $$ where ${\mathcal G}{x_i, \sigmai}, {\mathcal G}{x’_i, \sigma’_i}$ are two Gaussians. The variety of the LBP family comes from the choice of these Gaussians : they can have a fixed size but random positions (a la BRIEF), fixed sizes and positions (a la BRISK)1… The choice in FREAK was :

FREAK reconstruction as an inverse problem

In a previous post, I have briefly introduced our FREAK descriptor, which belongs to the more general family of the LBPs. In this post, I will state mathematically the problem of the reconstruction of an image patch given its descriptor, i.e. answering the question:

Can you guess which image part created this particular description vector ?

FREAK makes it into OpenCV trunk

Integrating our lab’s FREAK in your vision workflow is getting easier and easier!

Introducing FREAKs

OK, starting to mix posts in English and French since the English version of the blog is still buggy.