IncPCP-PTI

Description

This page focuses on the Panning and Translational Invariant Incremental Principal Component Pursuit (incPCP-PTI) algorithm. IncPCP-PTI is a newly developed fully incremental algorithm for video background modeling that extends incPCP-TI in order to cope with moving and panning cameras. The method continuously updates the low-rank component in order to align it to the current reference frame of the camera. To the best of our knowledge, this algorithm, named incPCP-PTI, is the first low rank plus additive matrix algorithm capable of handling both panning and jitter. 

In this page, we include the Matlab code and results on a synthetic panning and jitter video along with results on real videos from the CDNet 2014 dataset associated with the paper G. Chau and P. Rodriguez, "Panning and Jitter Invariant Incremental Principal Component Pursuit for Video Background Modeling" submitted to the 2nd "International Workshop on Robust Subspace Learning and Applications in Computer Vision" at the 2017 "International Conference on Computer Vision".

Matlab Code

[Download Code]

The synthetic panning and jitter (SPJ) test video can be downloaded from here. This dataset was generated from the Neovision dataset by selecting a subregion and translating and adding jitter to it. The other two test videos can be obtained from CDNet 2014 dataset webpage.  The code can be found at the bottom of this page. The examples here showed can be run using the following script.

myFlags = incAMFastPCPinputPars('TI_search');

myFlags.showFlag = 1;

myFlags.shrinkRule='l1proj';

myFlags.shrinkAlpha=0.75;

% ghost supression parameters

myFlags.ghost = 1;

myFlags.ghostUniModOff = 0.1;

myFlags.ghostFrames = 20;

myFlags.ghostFactor = 0;

% PTI parameters

myFlags.TI = 2; 

myFlags.TIextraOneLoopSolve = 1;

myFlags.folder_results = %folder to which to save the results

folder = % input folder

incrementalPCP_ball(folder,1, 3, 30, myFlags);

Results

Results for SPJ test video. Left: Original video. Right: Sparse component found with incPCP-PTI.

Results for the continuous pan video of the CDNet 2014 dataset. Left: Original video. Right: Sparse component found with incPCP-PTI.

Results for the intermittent pan video of the CDNet 2014 dataset. Left: Original video. Right: Sparse component found with incPCP-PTI.