300W Results

This jupyter notebook summarises the results of the active appearance models which were trained using the menpo Python Package. For more information regarding the training of these models please refer to:

Section 4 of the Review Paper
The Active Appearance Model base class contained within aam.py
The holisitc AAM and patch AAM models described within 300W_aam_hol.py and 300W_aam_patch.py

This notebook will contained the cumulative error distribution curves produced by the models as well as a selection from literature.

Resource Utilisation

Model	Time (hh:mm:ss)	Avg # CPUs Used	Memory Used (GB)
hAAM	00:15:41	5.12	4.4
pAAM	00:09:10	3.79	4.3

Test Set

A brief description of the test set used during the experiment. Note that the same test set was used for both holistic and patch experiments.



In [1]:

    
import os
with open('300W_aam_hol_testset.txt', 'r') as f:
    test_set_filenames = f.read().split('\n')

print("The test set comprised of %i images\n" % len(test_set_filenames))
print("Listing the first 10 images:")

# Print summary
test_set_filenames = [os.path.basename(filename) for filename in test_set_filenames]
print("\t".join(test_set_filenames[:10]) + "...")









    



The test set comprised of 90 images

Listing the first 10 images:
indoor_204.png	indoor_267.png	indoor_153.png	indoor_010.png	indoor_234.png	indoor_227.png	indoor_197.png	indoor_110.png	indoor_006.png	indoor_176.png...

Cumulative Error Distribution Curves

The following chart shows the normalised error distribution as produced by the holisitic and patch AAMs vs that described by Yan et al. in 2013¹, Zhou et al. in 2013² and He et al. in 2017³

Yan, J., Lei, Z., Yi, D., & Li, S. Z. (2013). Learn to combine multiple hypotheses for accurate face alignment. In Proceedings of the IEEE International Conference on Computer Vision (pp. 392–396). http://doi.org/10.1109/ICCVW.2013.126
Zhou, E., Fan, H., Cao, Z., Jiang, Y., & Yin, Q. (2013). Extensive facial landmark localization with coarse-to-fine convolutional network cascade. In Proceedings of the IEEE International Conference on Computer Vision (pp. 386–391). Institute of Electrical and Electronics Engineers Inc.
Milborrow, S., Bishop, T. E., & Nicolls, F. (2013). Multiview active shape models with SIFT descriptors for the 300-W face landmark challenge. In Proceedings of the IEEE International Conference on Computer Vision (pp. 378–385). https://doi.org/10.1109/ICCVW.2013.57



In [2]:

    
import numpy as np

# Import and configure matplotlib
import matplotlib
%matplotlib inline
matplotlib.rcParams['figure.figsize'] = (15, 10)
matplotlib.rcParams['font.size'] = 14
import matplotlib.pyplot as plt


# These files contains the CED data from experimentation & literature
FILES = ['300W_aam_hol.npy', '300W_aam_patch.npy', '300W_milb_indoor_outdoor.npy', 
         '300W_yan_2013_indoor_outdoor.npy', '300W_zhou_2013_indoor_outdoor.npy']
LABELS = ['hAAM', 'pAAM', 'Milborrow et al. 2013', 'Yan et al. 2013', 'Zhou et al. 2013']

for filename, label in zip(FILES, LABELS):
    
    # Load the data
    data = np.load(filename)
    
    # Normalise
    data[1] /= np.max(data[1])
    
    # Plot
    plt.plot(data[0], data[1], label=label)


# Add details to plot
plt.legend(loc='lower right', prop={'size': 13});
plt.grid();
plt.xlabel('Point-to-point error normalised by intra-ocular distance');
plt.ylabel('Proportion of Images');