Main MenpoFit Widgets

These are top-level functions for visualizig the various menpofit objects with high-level interactive widgets.

  1. Basics
  2. Train Models
  3. Holistic AAM and ATM Widgets
  4. Patch-based AAM and ATM Widgets
  5. CLM Widgets
  6. Fitting Result
  7. Cumulative Error Distributuon

We highly recommend that you render all matplotlib figures inline the Jupyter notebook for the best menpowidgets experience. This can be done by running
%matplotlib inline
in a cell. Note that you only have to run it once and not in every rendering cell.

1. Basics

Let's first import all the widgets:

from menpowidgets import (visualize_shape_model, visualize_appearance_model, visualize_aam,
                          visualize_atm, visualize_patch_appearance_model, visualize_patch_aam,
                          visualize_patch_atm, visualize_expert_ensemble, visualize_clm,
                          visualize_fitting_result, plot_ced, visualize_images)

They are all functions which have some common arguments:

  • style: It can be either 'coloured' or 'minimal'. The 'coloured' style uses a colouring theme that is different for each widget. The 'minimal' style is very simple with black and white colours.
  • figure_size: This argument is a tuple that defines the size of the rendered figure in inches. The figure size can also be controlled from the Renderer options within the widgets.
  • browser_style: It can be either 'buttons' or 'slider'. This argument exists in widgets that visualize a list of objects (e.g. shape models or fitting results). If 'buttons', then the object selection will be done by the </i> / buttons. If 'slider', then the object selection is done using a slider widget.

Note that all widgets can get as input a list of objects with totally different attributes between them. For example, list of results with different number of shapes or list of PCAModel with different number of parameters. Finally, they all have a Renderer tab that has many rendering-related options (such as lines, markers, axes, legend, grid, image) and an Export tab that allows the user to save the figure to file.

2. Train Models

Before moving on, let's train some models. First let's import what is needed.

%matplotlib inline
import menpo.io as mio
from menpo.feature import igo
from menpo.visualize import print_dynamic

from menpofit.aam import HolisticAAM, PatchAAM, LucasKanadeAAMFitter
from menpofit.atm import HolisticATM, PatchATM
from menpofit.clm import CLM, CorrelationFilterExpertEnsemble
from menpofit.fitter import noisy_shape_from_shape

Now let's load a few some training and fitting images from the LFPW database. Since the aim of this widget is to demonstrate the functionalities of widgets, we don't need to train the most performant models, thus only 60 training images are enough for fast training.

def load_images(path, crop_proportion=None, max_images=None, verbose=True):
    images = []
    for im in mio.import_images(path, max_images=max_images, verbose=verbose):
        # Crop image
        im = im.crop_to_landmarks_proportion(crop_proportion)
        # Congert to greyscale
        if im.n_channels == 3:
            im = im.as_greyscale()
        # Append image object
        images.append(im)
    return images


training_images = load_images('/home/nontas/Dropbox/lfpw/trainset/',
                              crop_proportion=0.2, max_images=60)
test_images = load_images('/home/nontas/Dropbox/lfpw/testset/',
                          crop_proportion=0.3, max_images=10)
training_shapes = [im.landmarks['PTS'].lms for im in training_images]
Found 60 assets, index the returned LazyList to import.
Found 10 assets, index the returned LazyList to import.

Let's visualize the training images:

visualize_images(training_images)

png

Let's train a simple holistic and patch-based AAM:

aam = HolisticAAM(training_images, holistic_features=igo, scales=(0.5, 1.0), diagonal=150,
                  max_shape_components=20, max_appearance_components=150, verbose=True)
- Computing reference shape                                                     Computing batch 0
- Building models
  - Scale 0: Done
  - Scale 1: Done


/home/nontas/Documents/Research/menpofit/menpofit/builder.py:341: MenpoFitModelBuilderWarning: The reference shape passed is not a TriMesh or subclass and therefore the reference frame (mask) will be calculated via a Delaunay triangulation. This may cause small triangles and thus suboptimal warps.
  MenpoFitModelBuilderWarning)
patch_aam = PatchAAM(training_images, holistic_features=igo, scales=(0.5, 1.), diagonal=170,
                     patch_shape=[(15, 15), (21, 21)], max_shape_components=20,
                     max_appearance_components=150, verbose=True)
- Computing reference shape                                                     Computing batch 0
- Building models
  - Scale 0: Done
  - Scale 1: Done

Let's also train a holistic and patch-based ATM using the second image as template:

atm = HolisticATM(training_images[1], training_shapes, holistic_features=igo, scales=(0.5, 1.),
                  diagonal=150, max_shape_components=20, verbose=True)
- Computing reference shape                                                     Computing batch 0
- Building models
  - Scale 0: Done
  - Scale 1: Done


/home/nontas/Documents/Research/menpofit/menpofit/builder.py:341: MenpoFitModelBuilderWarning: The reference shape passed is not a TriMesh or subclass and therefore the reference frame (mask) will be calculated via a Delaunay triangulation. This may cause small triangles and thus suboptimal warps.
  MenpoFitModelBuilderWarning)
patch_atm = PatchATM(training_images[1], training_shapes, holistic_features=igo, scales=(0.5, 1.),
                     diagonal=170, patch_shape=[(15, 15), (21, 21)], max_shape_components=20,
                     verbose=True)
- Computing reference shape                                                     Computing batch 0
- Building models
  - Scale 0: Done
  - Scale 1: Done

Finally, let's train a simple CLM model:

clm = CLM(training_images, diagonal=150, scales=[1.], holistic_features=igo, patch_shape=(19, 19),
          context_shape=(34, 34), expert_ensemble_cls=CorrelationFilterExpertEnsemble,
          max_shape_components=20, verbose=True)
- Computing reference shape                                                     Computing batch 0
- Training models
  - Done

3. Holistic AAM and ATM Widgets

The shape models and appearance models of the Holistic AAM can be visualised as:

aam.view_shape_models_widget()

png

aam.view_appearance_models_widget()

png

Hit the button to view an animation of the variance of each component. The AAM can be viewed as:

aam.view_aam_widget()

png

Note that the above class functions are equivalen to:

visualize_shape_model(aam.shape_models)
visualize_appearance_model(aam.appearance_models)
visualize_aam(aam)

Now let's also visualize the trained holistic ATM.

atm.view_atm_widget()

png

4. Patch-based AAM and ATM Widgets

Similarly to the holistic case, the patch-based AAM can be visualised as follows:

patch_aam.view_appearance_models_widget()

png

patch_aam.view_aam_widget()

png

and the patch-based ATM:

patch_atm.view_atm_widget()

png

5. CLM Widgets

Let's also visualize the CLM widgets:

clm.view_expert_ensemble_widget()

png

clm.view_clm_widget()

png

6. Fitting Result

Let's fit the patch-based AAM on the test images in order to create a list of iterative fitting results.

fitter = LucasKanadeAAMFitter(patch_aam, n_shape=[3, 12], n_appearance=150)
fitting_results = []
for j, im in enumerate(test_images):
    print_dynamic('{}/{}'.format(j + 1, len(test_images)))
    groundtruth_shape = im.landmarks['PTS'].lms
    initial_shape = noisy_shape_from_shape(fitter.reference_shape, groundtruth_shape,
                                           noise_type='uniform', noise_percentage=0.02,
                                           allow_alignment_rotation=False)
    fr = fitter.fit_from_shape(im, initial_shape, gt_shape=groundtruth_shape,
                               max_iters=[10, 15])
    fitting_results.append(fr)
10/10                                                                           

The fitting results can be visualized as:

visualize_fitting_result(fitting_results)

png

Of course, a result has a built-in view_widget() function

fitting_results[0].view_widget()

png

Let's now create a list of Result objects with different properties and visualize them:

frs = [fitting_results[0]]
frs.append(fitting_results[1].to_result(pass_image=False, pass_initial_shape=False, pass_gt_shape=False))
frs.append(fitting_results[2].to_result(pass_image=True, pass_initial_shape=False, pass_gt_shape=False))
frs.append(fitting_results[3].to_result(pass_image=False, pass_initial_shape=True, pass_gt_shape=False))
frs.append(fitting_results[4].to_result(pass_image=False, pass_initial_shape=False, pass_gt_shape=True))
frs.append(fitting_results[5].to_result(pass_image=True, pass_initial_shape=True, pass_gt_shape=False))
frs.append(fitting_results[6].to_result(pass_image=True, pass_initial_shape=False, pass_gt_shape=True))
frs.append(fitting_results[7].to_result(pass_image=False, pass_initial_shape=True, pass_gt_shape=True))
frs.append(fitting_results[8].to_result(pass_image=True, pass_initial_shape=True, pass_gt_shape=True))
frs.append(fitting_results[9])

visualize_fitting_result(frs)

png

Note that the widget adapts itself to the properties of each fitting result object.

7. Cumulative Error Distributuon

The fitting errors can be nicely visualized with a widget for plotting Cumulative Error Distributions (CED). Let's first create a list with the initial and final fitting errors.

initial_errors = [r.initial_error() for r in fitting_results]
final_errors = [r.final_error() for r in fitting_results]

and plot the curves:

plot_ced([initial_errors, final_errors], legend_entries=['Initial', 'Final'])

png

results matching ""

    No results matching ""