tnarayanan / CRUMPL

Winning Stanford CS231n project code that uncrumples images using adversarial paradigms.

Sixteen Pixels is (Almost) All You Need: Crafting Parameterized Image Uncrumpling Models

Stanford CS231n final project, Spring 2022

Authors (alphabetical by last name):

• Maximilian Du
• Niveditha Iyer
• Tejas Narayanan

Installing dataset

All data should be installed under the data subdirectory.

We are using the validation subset of ImageNet as our full dataset, since the full ImageNet dataset is extremely large.

To access ImageNet data, sign up for an account at https://image-net.org/. Then, download the "blurred validation images" under the "Face obfuscation in ILSVRC" heading. This will download val_blurred.gz. Extract this archive into the data/val_blurred directory. The structure should look like the following:

CRUMPL
│  ...
└─ data
   └─ val_blurred
      └─ n01440764
      └─ n01443537
      └─ ...

Generating data

Download Blender from https://www.blender.org/download/. Then, open paper_gen.blend and navigate to the Scripting tab at the top. In the console that appears on the bottom left, enter the following two lines to install tqdm, a progress bar library:

>>> import pip
>>> pip.main(['install', 'tqdm'])

Replace the constants DATA_PATH and EXPORT_PATH based on your computer's file path. Finally, run the file by pressing the run button on the menu bar of the code editor area.