## Finds in arxiv, January.

Repost from googleplus stream

Finds in arxiv

**Computer vision**

*Convex Relaxations of SE(2) and SE(3) for Visual Pose Estimation*

Finding camera position from the series of 2D (or depth) images is one of the most common (and difficult) task of computer vision.

The biggest problem here is incorporating rotation into cost(energy) function, that should minimize reprojection(fro 2D) error (seehttp://en.wikipedia.org/wiki/Bundle_adjustment). Usually it’s done by minimization on manifold – locally presenting rotation parameter space as linear subspace. The problem with that approach is that initial approximation should be good enough and minimization goes by small steps pojecting/reprojecting on manifold (SO(3) in the case) which can stuck in local minimum. Where to start if the is no initial approximation? Usually it’s just brute-forced with several initialization attempt. Here in the paper approach is different – no initial approximation needed, solution is straightforward and minimum is global. The idea is instead of minimizing rotation on the sphere minimize it inside the ball. In fact it’s a classical **convexification** approach – increase dimentionality of the problem to make it convex. The pay is a lot higher computational cost of cause

http://arxiv.org/abs/1401.3700

*A Sparse Outliers Iterative Removal Algorithm to Model the Background in the Video Sequences*

Background removal without nuclear norm. Some “good” frames chosen as dictionary and backround represented in it.

http://arxiv.org/abs/1401.6013

**Deep Learning**

*Learning Mid-Level Features and Modeling Neuron Selectivity for Image Classification*

Mid level features is a relatively new concept but a very old practice. It’s a building “efficient” subset of features from the set of features obtained by low-level feature descriptor(like SIFT, SURF, FREAK etc) Before it was done by PCA, later by sparse coding.

What authors built is some kind of hibrid of convolutional network with dictionary learning. Mid level features fed into neuron layer for sparsification and result into classifier. There are some benchmarks but no CIFAR or MINST

http://arxiv.org/abs/1401.5535

**Optimization and CS**

*Alternating projections and coupling slope*

Finding intersection of two non-convex sets with alternating projection. I didn’t new that transversality condition used a lot in convex geometry too.

http://arxiv.org/abs/1401.7569

*Bayesian Pursuit Algorithms*

Bayesian version of hard thresholding. Bayesian approach is in how threshold is chosen and update scaled.

http://arxiv.org/abs/1401.7538