15
Shoe Classification
Multi-class CNN for shoe image classification. Error analysis and visualization of the features learned by the network.
Accuracy: 66%
Stack
What happened?
- 01
The question
Can a CNN distinguish shoe types — beyond the simple “it's a shoe”?
- 02
The problem
Multi-class is another world: close classes (boots vs high-top sneakers) and a real risk of overfitting.
- 03
The hypothesis
A multi-class CNN with learned-feature visualization will reveal what the network is actually looking at.
- 04
The build
Multi-class CNN for shoe classification: training, error analysis, and visualization of learned features.
- 05
The failure
66% accuracy over several classes: the errors showed the network focusing on color and background rather than shape.
- 06
The discovery
Visualizing learned features is opening the black box. The network was looking at the wrong things, and data augmentation helped it look at the shoe.
- 07
The result
A 66%-accuracy multi-class CNN, with error analysis and learned-feature visualization.
- 08
The next frontier
Transfer learning, more varied data, and segmentation to isolate the shoe from the background.