Mini Places Challenge

MIT 6.819 Mini Places Challenge by Maryam Archie and Sandeep Silwal.

Getting Started

1. To run these scripts, you first need to download the image data. Be warned - it is going to take a long time to download.

2. Edit the data paths in augment_images.py and vgg16.py.

3. Download and Python 3 (if you don't already - preferably from Anaconda)

4. Install the following dependencies:

pip install tensorflow
pip install tensorflow-gpu
pip install scipy
pip install pillow

5. It is strongly recommended that you use a GPU when running vgg16.py.

Augmenting the image data set

With augment_images.py, you can:

• Increase the dataset from 100,000 to 400,000 images with label-preserving transformations
• Update data\train.txt with the new images
• Calculate the new mean of the data set

For your convenience, this has already been done and is available in the mini-places-data repository. In addition, the data mean in vgg16.py has been updated to reflect this.

References

• Original Databases: Places2 Database, Places1 Database
• Mini Places Challege