This project was done as a part of the Identify Characers From Product Images competition. The competition was hosted in CrowdAnalytix.
The data consisted of product images like t-shirts, bags, keychains, mobile covers, etc. with characters graphics.
Training Set: 6694 images across 42 categories Test Set: 3727 images
The data had label inconsistency of 2-3%. Had to manually resolve the label inconsitency.
As the data provided by CrowdAnalytix was not equally distributed across the 42. Some categories had fewer data comparatively to the other categories. To resolve this data insufficiency among the categories we downloaded the additional data with the help of gi2ds and this tutorial created by Andrian Rosebrook .
You can download the data from here (Comprises the data provided by CrowdAnalyticsX and the above mentioned additional data). The filenames of images from Crowdanalytix starts with Cax_train and the other images filenames start with number.
The following were the 42 categories for classification.
- Angry Birds
- Baloo
- Bart simpson
- Ben
- Bulbasaur
- Charizard
- Charlie brown
- Charmender
- Chicken_little
- Cinderella
- Darth_vader
- Disney_princes
- Donald_duck
- Godzilla
- Goku
- Goofy
- Han-solo
- Harry_potter
- Hellokitty
- Itachi
- John_Cena
- Jojosiwa
- Kakashi
- Marilyn_monroe
- Mickey_mouse
- Minions
- Naruto
- Pikachu
- Pokemon
- Popeye
- Power_rangers
- R2-D2
- Roman_reigns
- Scoopy Doo
- SpongeBob SquarePants
- Squirtle
- Teenage_mutant_ninja_turtles
- Tom and Jerry
- Toy_story_characters
- Vampirina
- Vegeta
- Winnie the poo
Pair of confused categories with minimum value of 2
[('Cinderella', 'disney_princes', 6), ('pokemon', 'pikachu', 6), ('Baloo', 'Godzilla', 5), ('Charlie brown', 'goofy', 5), ('Charlie brown', 'Bart simpson', 4), ('popeye', 'power_rangers', 4), ('Roman Reigns', 'han-solo', 3), ('Scoopy Doo', 'Bart simpson', 3), ('disney_princes', 'Baloo', 3), ('disney_princes', 'Cinderella', 3), ('donald_duck', 'goofy', 3), ('goofy', 'mickey_mouse', 3), ('popeye', 'goofy', 3), ('vampirina', 'jojosiwa', 3), ('Bart simpson', 'Charlie brown', 2), ('Charlie brown', 'teenage_mutant', 2), ('Chicken_little', 'power_rangers', 2), ('Godzilla', 'Tom and Jerry', 2), ('Godzilla', 'goofy', 2), ('Goku_1', 'goofy', 2), ('Goku_1', 'pokemon', 2), ('John Cena', 'Roman Reigns', 2), ('Tom and Jerry', 'goofy', 2), ('charizard', 'Goku_1', 2), ('disney_princes', 'Bart simpson', 2), ('harry_potter', 'han-solo', 2), ('itachi', 'Goku_1', 2), ('kakashi', 'power_rangers', 2), ('naruto', 'Goku_1', 2), ('naruto', 'pokemon', 2), ('pikachu', 'pokemon', 2), ('pokemon', 'angrybirds', 2), ('popeye', 'mickey_mouse', 2), ('squirtle', 'bulbasaur', 2), ('squirtle', 'power_rangers', 2), ('squirtle', 'teenage_mutant', 2), ('vegeta', 'goofy', 2)]
The validation accuracy obtained was 90.7%
The public leaderboard ranking: 12th with test accuracy of 89%
The private leaderboard ranking: 11th with test accuracy of 90%
fast.ai
DL Techniques Used From fastai course