Multi-view face detection trained on AFLW dataset.

This repo has the auxiliary Matlab and C++ code in order to replicate the face detection experiments in our paper. If you use this code for your own research, you must reference our journal paper:

• BAdaCost: Multi-class Boosting with Costs. Antonio Fernández-Baldera, José M. Buenaposada, and Luis Baumela. Pattern Recognition, Elsevier. In press, 2018. DOI:10.1016/j.patcog.2018.02.022

Requirements

• Clone toolbox.badacost.public repo, with our modified version of Piotr Dollar toolbox with the BAdaCost algorithm with cost-sensitive trees. Go to its directory and execute Matlab. Then from Matlab prompt, execute addpath(createpath(PATH_TO_TOOLBOX)) and then toolboxCompile.

• Clone toolbox.badacost.faces.public, a set of tools and code to perform the faces detection experiments with the AFLW, AFW, PASCAL faces and FDDB datasets.

• We train our detector with face images from the Annotated Facial Landmarks in the Wild dataset. Thus, from the AFLW site you have to download:

  • aflw-images-0.tar.gz,
  • aflw-images-2.tar.gz and
  • aflw-images-3.tar.gz

  by decompressing the tar.gz files we will get the following directory structure:

    aflw
    `-- data
    	`-- flickr
    	   |-- 0
    	   |-- 2
    	   `-- 3
    
  

  The path to the flicker folder in the example, with the AFLW training images (dirs 0, 2, 3) will be refered as AFLW_PATH from now on.

• The negative face images are obtained from the PASCAL VOC 2007 dataset. Therefore we need to dowload the following files from the PASCAL VOC site:

  • VOCtest_06-Nov-2007.tar
  • VOCtestnoimgs_06-Nov-2007.tar
  • VOCtrainval_06-Nov-2007.tar

  by decompressing the tar files we will get the following directory structure:

      pascal_voc
    	`-- 2007
    	    `-- VOCdevkit
    		`-- VOC2007
    		   |-- Annotations
    		   `-- JPEGImages
  

  The path to the pascal_voc/2007/VOCdevkit/VOC2007 folder in the example, with the JPEGImages and Annotations subdirs will be refered as PASCAL_VOC2007_PATH from now on.

• The PASCAL Faces dataset, is extracted from different face images in the PASCAL VOC datasets. Therefore we need to dowload the following files from the PASCAL VOC site:

  • To be uncompressed in directory pascal_voc/2007:
    • VOCtrainval_06-Nov-2007.tar
    • VOCdevkit_08-Jun-2007.tar
    • VOCtest_06-Nov-2007.tar
    • VOCtestnoimgs_06-Nov-2007.tar
  • To be uncompressed in directory pascal_voc/2008:
    • VOCtrainval_14-Jul-2008.tar
    • VOCdevkit_14-Apr-2008.tar
    • VOC2008_test.tar
  • To be uncompressed in directory pascal_voc/2009_test:
    • InWGD4LN.tar
  • To be uncompressed in directory pascal_voc/2010_test:
    • VOC2010_test.tar
  • Untar the following files in the directory PASCAL_VOC/2011_test and then move 2011_test/Test/VOCdevkit to 2011_test/VOCdevkit
    • VO2011_test.tar
  • Untar the following files in the directory PASCAL_VOC/2012:
    • VOCdevkit_18-May-2011.tar
    • VOCtrainval_11-May-2012.tar

  by decompressing the tar files we will get the following directory structure:

     pascal_voc
    |
    |-- 2007
    |   `-- VOCdevkit
    |       `-- VOC2007
    |          |-- Annotations
    |          `-- JPEGImages
    |-- 2008_test
    |   `-- VOCdevkit
    |       `-- VOC2008
    |           |-- Annotations
    |           `-- JPEGImages
    |-- 2009_test
    |   `-- VOCdevkit
    |       `-- VOC2009
    |           |-- Annotations
    |           `-- JPEGImages
    |-- 2010_test
    |   `-- VOCdevkit
    |       `-- VOC2010
    |           |-- Annotations
    |           `-- JPEGImages
    |-- 2011_test
    |   `-- VOCdevkit
    |       `-- VOC2011
    |           |-- Annotations
    |           `-- JPEGImages
    `-- 2012
        `-- VOCdevkit
    	`-- VOC2012
    	    |-- Annotations
    	    `-- JPEGImages
  

  The path to the pascal_voc folder in the example, with the JPEGImages and Annotations subdirs will be refered as PASCAL_VOC_PATH from now on.

• The AFW dataset, which is in the file AFW.zip, can be downloaded from the AFW site. Decompress all the images in the .zip in a single directory and se the variable AFW_PATH in the main.m script to this directory.

• Download the FDDB dataset and prepare data for P.Dollar toolbox by running the script:

  fddb2PiotrFormatMulticlass.m