This is the official repository for B. Van Hoorick and C. Vondrick, "Dissecting Image Crops," ICCV 2021. In short, we investigate what traces are left behind by visual cropping.
Step 1: Populate data/train, data/val, and data/test with high-resolution image files; a constant aspect ratio is strongly preferred.
Step 2: Investigate the command line flags in train.py, and run python train.py with the desired arguments. This will instantiate a new training run with PyTorch checkpoint files in checkpoints/, and TensorBoard log files in logs/.
Step 3: Run python test.py --model_path /path/to/above/checkpoint/folder with relevant arguments to run the model on the test set.
In our project, we scraped Flickr based on this script by Sam Lavigne, using each line in google-10000-english-no-swears.txt (see this repository for more info) as search queries. We filtered the photos by an aspect ratio of 1.5, which is the most common value, resulting in a dataset of around 700,000 images. They were captured by diverse (but mostly high-end) camera brands, models, and pipelines.
There is a stubborn memory leak that builds up as you train over many epochs. I have tried many things but do not know how to prevent it.
@article{van2020dissecting,
title={Dissecting Image Crops},
author={Van Hoorick, Basile and Vondrick, Carl},
journal={ICCV 2021},
year={2020}
}