Landmarks recognition

This repository contains Python scripts and notebooks to generate classification models for 25 landmarks.

Five notebooks are present in the notebooks/ folder:

• 1 - Basic CNN which creates a model from a basic CNN ;
• 2 - Data Augmentation which used the same basic CNN with data augmentation ;
• 3 - Bottleneck features which uses the bottleneck features from a pre-trained VGG16 CNN ;
• 4 - Fine-tuning which fine-tunes the last convolutional block of a pre-trained VGG16 CNN ;
• 5 - Adversarial examples with FGSM which generates adversarial examples with FGSM (Fast Gradient Sign Method) ;
• 6 - Cleverhans benchmark which bencharmks our basic CNN model with cleverhans package using FGSM attack.

The five notebooks were run on a p2.xlarge AWS EC2 instance using the Deep Learning AMI.

Local development

First, make sure you have Python 3 installed on your machine, along with the following packages :

• tqdm
• Tensorflow
• Keras
• cleverhans

The dataset comes from here. A Python script is provided to download all the images:

$ python3 utils/download_data.py 'train'
$ python3 utils/download_data.py 'validation'

It can take several hours depending on your internet connection speed.

Once all the photos are in the data folder, run a cleanup script in order to remove not found images :

$ python utils/dataset_cleanup.py

You can now execute the notebooks on your local environment !