Support and Weekly deliverables

Instructions for students:

• Issues: Open new issues in this repository to pose your questions.
• Deliverables: A new issue will be created in this repository for you to upload your weekly deliverables (Google Slides). Find it and answer with a public link to your slides before the deadline.

Code Instructions

Requirements

• Python 2.7
• Install pip following these instructions.
• Run pip install -r requirements.txt to install code dependencies.
• Install OpenCV 2.4.x from source (instructions here) or using the package manager in Canopy with the free academic subscription (the latter is recommended for simplicity).

Note: version 3.x of OpenCV will not work with the current version of the code since SIFT and SURF extraction modules were removed.

Setup

• Edit the file src/params.py so that params['root'] points to the directory where you have the dataset TerrassaBuildings900 and where you would like to store all the intermediate files. You can also modify some of the parameters there (such as the codebook size, the keypoint detector, etc.)
• Run src/params.py. Notice that this will create directories as well, so make sure you did the previous step !
• Run src/build_database.py to generate text files with image IDs for both the training and validation sets.

Feature extraction

• Run src/get_features.py. This will first train a codebook, and then build BoW vectors for the training and validation/test images.

S5: Ranking Instructions

• Run src/rank.py to compute distances from validation images to train images and generate the rankings.
• Run src/eval_rankings.py to obtain the mean Average Precision of your rankings.
• For a step by step tutorial and a better analysis of the results, check notebooks/gdsa_s5.ipynb

S6: Classification Instructions

• Run src/classify.py. This will train a classifier and use it to predict classes for validation images.
• Run src/eval_classification.py to get evaluation metrics of your classifier predictions.
• Check notebooks/gdsa_s6.ipynb to see a tested example.

Kaggle rankings

There is a script utils/save_for_kaggle.py that will take your rankings and convert them to be uploaded to Kaggle.

Deep features (coming soon !)

The script src/deepfeats.py extracts VGG-16 features using Keras and generates rankings in Kaggle format. It assumes you have first run src/build_database.py. This is an alternative to the feature extraction pipeline that relies on local features (e.g. SIFT)