Merge branch 'statistics' into 'master'

Add a script for analyzing database face sizes

See merge request !62

bob/pad/face/script/statistics.py

+"""Gets statistics on the average face size in a video database.
+"""
+import logging
+import click
+import numpy as np
+from os.path import expanduser
+from bob.extension.scripts.click_helper import (
+    verbosity_option, ConfigCommand, ResourceOption)
+from bob.ip.facedetect import (
+    bounding_box_from_annotation, expected_eye_positions, BoundingBox)
+
+logger = logging.getLogger(__name__)
+BF = 'Bona Fide'
+PA = 'Presentation Attack'
+
+
+@click.command(entry_point_group='bob.bio.config', cls=ConfigCommand)
+@click.option('--database', '-d', required=True, cls=ResourceOption,
+              entry_point_group='bob.pad.database')
+@click.option('-o', '--output', default='face_sizes.png')
+@click.option('--database-directories-file', cls=ResourceOption,
+              default=expanduser('~/.bob_bio_databases.txt'))
+@verbosity_option(cls=ResourceOption)
+def statistics(database, output, database_directories_file, **kwargs):
+    """Statistics on face size in video databases.
+
+    \b
+    Parameters
+    ----------
+    database : :any:`bob.pad.database`
+        The database that you want to annotate. Can be a ``bob.pad.database``
+        entry point or a path to a Python file which contains a variable
+        named `database`.
+    output : str
+        Path to the saved figure.
+    database_directories_file : str
+        Path to a custom ``~/.bob_bio_databases.txt`` file.
+    verbose : int, optional
+        Increases verbosity (see help for --verbose).
+
+    \b
+    [CONFIG]...            Configuration files. It is possible to pass one or
+                           several Python files (or names of ``bob.bio.config``
+                           entry points) which contain the parameters listed
+                           above as Python variables. The options through the
+                           command-line (see below) will override the values of
+                           configuration files.
+    """
+    logger.debug('database: %s', database)
+    logger.debug('output: %s', output)
+    logger.debug('database_directories_file: %s', database_directories_file)
+    logger.debug('kwargs: %s', kwargs)
+
+    # Some databases need their original_directory to be replaced
+    database.replace_directories(database_directories_file)
+
+    biofiles = database.objects(groups=None, protocol=database.protocol)
+    biofiles = sorted(biofiles)
+
+    logger.info("Gathering face size statistics of %d samples", len(biofiles))
+
+    face_sizes_dict = {BF: [], PA: []}
+    for i, biofile in enumerate(biofiles):
+        for annot in database.annotations(biofile).values():
+            # get the bounding box
+            for source in ('direct', 'eyes', None):
+                try:
+                    bbx = bounding_box_from_annotation(source=source, **annot)
+                    break
+                except Exception:
+                    if source is None:
+                        raise
+                    else:
+                        pass
+            # record face size
+            if biofile.attack_type is None:
+                face_sizes_dict[BF].append(bbx.size)
+            else:
+                face_sizes_dict[PA].append(bbx.size)
+
+    if output:
+        import matplotlib.pyplot as plt
+        # from matplotlib.backends.backend_pdf import PdfPages
+        # pp = PdfPages(output)
+
+    for attack_type, face_sizes in face_sizes_dict.items():
+        click.echo(attack_type)
+        face_sizes = np.array(face_sizes)
+        # get statistics on the face sizes
+        for name, array in (
+            ('Height', face_sizes[:, 0]),
+            ('Width', face_sizes[:, 1]),
+        ):
+            click.echo('min: {}, mean: {}, max: {}, std: {:.1f} for {}'.format(
+                array.min(), int(array.mean()), array.max(), array.std(),
+                name))
+        # print the average eye distance assuming bounding boxes are from
+        # bob.ip.facedetect or the annotations had eye locations in them
+        bbx = BoundingBox((0, 0), face_sizes.mean(axis=0))
+        annot = expected_eye_positions(bbx)
+        eye_distance = np.linalg.norm(
+            np.array(annot['reye']) - np.array(annot['leye']))
+        click.echo('Average eye locations: {}'.format(annot))
+        click.echo('Average eye distance: {}'.format(int(eye_distance)))
+
+        if not output:
+            continue
+
+        # plot the face sizes
+
+        # plt.hist2d(face_sizes[:, 1], face_sizes[:, 0], bins=500)
+        # plt.xlabel('Width')
+        # plt.ylabel('Height')
+        # plt.grid()
+
+        # from matplotlib import cm
+        # from mpl_toolkits.mplot3d import Axes3D
+        # Z, xedges, yedges, _ = plt.hist2d(
+        #     face_sizes[:, 1], face_sizes[:, 0], bins=500, normed=True)
+        # xcenters = (xedges[:-1] + xedges[1:]) / 2
+        # ycenters = (yedges[:-1] + yedges[1:]) / 2
+        # X, Y = np.meshgrid(xcenters, ycenters)
+        # fig = plt.figure()
+        # ax = Axes3D(fig)
+        # ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=cm.viridis)
+
+        plt.hist(face_sizes[:, 1], density=True,
+                 bins='auto', label=attack_type,
+                 alpha=0.5)
+    if output:
+        plt.xlabel('Width of faces')
+        plt.ylabel('Probability Density')
+
+        plt.tight_layout()
+        plt.legend()
+        plt.savefig(output)
+        # pp.savefig(plt.gcf())

setup.py

@@ -129,6 +129,11 @@ setup(
             'idiap-user-machines = bob.pad.face.config.grid:idiap_user_machines',
             'small = bob.pad.face.config.grid:small',
         ],
+
+        # registered ``bob pad ...`` commands
+        'bob.pad.cli': [
+            'statistics        = bob.pad.face.script.statistics:statistics',
+        ],
     },
 
     # Classifiers are important if you plan to distribute this package through