Code for Robust Kronecker-Decomposable Component Analysis for Low-Rank Modeling (ICCV 2017).

Author: Mehdi Bahri

This code has been adapted from my MSc thesis work at Imperial College (summer of 2016).

Python implementation: https://github.com/mbahri/RKCA.

• Matlab (tested on Matlab 2015a and later)
• Image Processing Toolbox for noise generation (imnoise)
• Control System Toolbox for discrete-time Silverster's equation solver (dlyap)
• Optional: Parallel Computing Toolbox

The main function is kdrsdl.

function [ Data, Info ] = kdrsdl( X, varargin)

Data and Info are structures that contain, respectively, the recovered components of the decomposition, and some information to monitor convergence (number of iterations and value of the stopping criterion at each step).

The Experiments folder contains sample data sets along with helper functions.

Reference of the supported options.

Upper-bound on the mode-1 and mode-2 ranks, the maximum dimension of the code matrices Rn is r*r. Defaults to min(m, n) for m*n matrices.

The lambda parameter in the optimization problem. Defaults to 1 / sqrt(N * max(m, n)) for m*n*N tensors.

Tolerance for convergence, the algorithm will stop when the stopping criterion is < tol. Defaults to 1e-7.

Maximum number of iterations. Defaults to 150.

The multiplicative factor by which the dual-steps mu and mu_k are increased. Defaults to 1.2

If true, the algorithm will stop as soon as the reconstruction error increases (regardless of weather it would have later decreased). Defaults to false

Allows giving a weight different from 1 to ||A||_F^2 in the optimization problem. Defaults to 1.

Allows giving a weight different from 1 to ||B||_F^2 in the optimization problem. Defaults to 1.

The alpha parameter in the optimization problem. Defaults to 1e-2.

Allows overriding the initialization of mu.

Allows feeding in the ground-truth low-rank tensor (useful to write visualizations to monitor training).

Allows feeding in the ground-truth sparse tensor (useful to write visualizations to monitor training).

Robustly estimate the sample mean (see last section of this readme). Defaults to false

Run some of the updates in parallel with the Matlab Parallel Computing Toolbox. Used for computations that are identical and independent on all frontal slices of a tensor. Default to false (warning: enabling it may actually lead to decreased performance due to overhead).

Level of details with which to display timing information.

• 0: No timing
• 1: Display time taken by each iteration
• 2: Add time spent in intialization to 1
• 3: Add time spent updating the codes R to 2, notifies when A, B, E have been updated

We provide both the Yale-B and Facade benchmarks, the procedure is the same in both cases:

% Load the original and noisy image(s)
% [Original, Noisy] = yale_sp(rescaling_factor, noise_level)
% Call facade_sp for the Facade benchmark

[O, X] = yale_sp(1, 0.3);

% See parameters section for the supported options
[Data, Info] = kdrsdl(X);

% Display reconstruction
imshow(Data.L(:,:,1), []);

We provide the Hall data set and two ground-truth frames.

% Load the frames and the ground-truth
load_hall
% There are now three variables:
% * O: The frames
% * GT: The ground-truth frames
% * GT_frames: The ids of the two frames that match the ground truth

% See below for explanations about the mean parameter
[Data, Info] = kdrsdl(O, 'mean', 1)

For background subtraction experiments, the implementation differs slightly from the one presented in the paper in that the mean value of the low-rank tensor's frontal slices is estimated separately and robustly as in [1]. That is, we write and, at step , we compute the estimator:

We then subtract the current estimated mean, and the outliers, to the input tensor. Line 5 of Algorithm 1 becomes:

All other steps remain identical.

[1] G. Mateos and G. B. Giannakis. Robust PCA as Bilinear 937 Decomposition With Outlier-Sparsity Regularization. IEEE 938 Transactions on Signal Processing, 60(10):5176–5190, 10 939 2012