Task : Classification of 196 classes of cars with less than 9k images for training. It's intend for replication of the work :
Monza: Image Classification of Vehicle Make and Model Using Convolutional Neural Networks and Transfer Learning
- use pre-trained googlenet
- train with normal rates
- Achieved 87% top1 accuracy and 97% top5 acc with 10000 iterations
Lessons learned :
1. Training from scratch is very hard with less data and more class.
2. Bounding box of the car really helps
3. fine tuning need to select a good lr and make the bottom layers learn slow, but the last few layers learn fast.