The following data augmentation has been applied to increase the no of images in the training set
Flip horizontal
Lighting
Zooming
Warping
Confusion Matrix
Pair of confused categories with minimum value of 2
[('motogp', 'formula1', 5),
('badminton', 'tennis', 4),
('weight_lifting', 'wrestling', 3),
('wrestling', 'wwe', 3),
('badminton', 'table_tennis', 2),
('basketball', 'volleyball', 2),
('boxing', 'wrestling', 2),
('hockey', 'ice_hockey', 2),
('kabaddi', 'hockey', 2),
('tennis', 'table_tennis', 2),
('weight_lifting', 'wwe', 2),
('wrestling', 'kabaddi', 2)]
Sample Images from the confused categories:
For some images you can see the prediction is correct but the manual
labelling was done wrong. This shows the CNN are less susceptible to
human mistakes.