KCCA-ReId

NEWS: We improved our method! Please, check it out MCK-CCA web page

We provided the implementation of the paper Giuseppe Lisanti , Iacopo Masi , Alberto Del Bimbo, "Matching People across Camera Views using Kernel Canonical Correlation Analysis”, Eighth ACM/IEEE International Conference on Distributed Smart Cameras, 2014.

##Requirements You need the following software to run the code:

Aug. 2016: It seems that the link to the KCCA package from Hardoon is not publicly available anymore on his website. If you want to run our approach your have to manually request his code dropping an email to Dr. Hardoon.

~~Sept. 2014: Please, note that the code is automatically attempting to download third-party libraries and data.~~

Demo Example

To run our code just run demo_reid_kcca.m

You can change the dataset and enable CCA comparison with the following parameters:

datasetname='VIPeR'; %VIPeR %PRID
ccaON = 0; %% can be turned off since it is slow.

and you will get something like this:

>Computing Trial 1...
>Applying Kernel to Train and Test...
>Computing KCCA  on the training set...
Centering Kx and Ky
Decomposing Kernel with PGSO
Computing nbeta from nalpha
>Projecting the test data...
>Computing distances...
>Evaluating results...
>Computing Trial 2...

Person Representation (Improved WHOS descriptor)

For representing a person we improved the WHOS descriptor presented in [2]. For the first part of the descriptor (color histograms), please see the details in [2]. (The only difference from [2] is that here we used a non-isotropic Gaussian kernel). Please see a video animation here.

For the second part of the descriptor (HOG + LBP) you can find here some details to re-implement the feature:

HOG: to compute HOG descriptor we quantized the gradients in 4 bins, pooling 8 px per cell, using block 2x2, no 360 degree, and cutting at 1 for normalization. We computed this in the window parameter specified below.

LBP: to compute LBP descriptor, we used the window parameter specified below, quantizing LBP codes using vl_feat code, building an LBP histogram for each patch 16x16 px (thus cellSize 16px).

Note that both the descriptors are computed on a specific window of the person image (resized to 128x64). This window corresponds to remove 8 px for each side and so computing the descriptor in a window that is [9:120,9:56].

Changelog

1.2 Aug. 2016 - Added pointer to new Hardoon's website for KCCA package.
1.1 Jan. 2015 - Added more information on WHOS descriptor
1.0 Sept. 2014 - Inital Release

Citation

Please cite our paper with the following bibtex if you use our code:

@article{lisanti:icdsc14,
author = {Lisanti, Giuseppe and Masi, Iacopo and {Del Bimbo}, Alberto},
title = {Matching People across Camera Views using Kernel Canonical Correlation Analysis},
booktitle = {Eighth ACM/IEEE International Conference on Distributed Smart Cameras},
year = {2014}, }

and Hardoon's paper:

@article{hardoon:cca,
author = {Hardoon, David and Szedmak, Sandor and {Shawe-Taylor}, John},
title = {Canonical Correlation Analysis: An Overview with Application to Learning Methods},
booktitle = {Neural Computation},
volume = {Volume 16 (12)},
pages = {2639--2664},
year = {2004}, }

References

[1] Giuseppe Lisanti , Iacopo Masi , Alberto Del Bimbo, Matching People across Camera Views using Kernel Canonical Correlation Analysis”, Eighth ACM/IEEE International Conference on Distributed Smart Cameras, 2014.

[2] G. Lisanti, I. Masi, A. D. Bagdanov, and A. Del Bimbo, "Person Re-identification by Iterative Re-weighted Sparse Ranking", IEEE Transactions on Pattern Analysis and Machine Intelligence 2014.

sshan-zhao / KCCAReId