group-wise-iccv19

Pytorch implementation of: "Group-Wise Deep Object Co-Segmentation With Co-Attention Recurrent Neural Network" - Li et al. of ICCV 2019

This code is just a POC and overfits a single group of images. Please adjust it to fit your needs.

Run on Colab

Here's a colab for you.

Run locally

The best way to run in three steps is with Docker.

1. git clone https://github.com/francesco-p/group-wise-iccv19.git && cd group-wise-iccv19
2. sudo docker run --rm --gpus all -it -v $(pwd):/current ufoym/deepo bash
3. cd /current && python main.py

You don't have Docker? There I produced a requiremets.txt but I didn't test it.

Folder structure

.
├1─ data
│   └── 042_reproducible
│       ├── ground_truth
│       │   ├── 369293688_87888a7a6c.png
│       │   └── ...
│       └── images
│           ├── 369294302_18dca2ab5b.jpg
│           └── ...
├── LICENSE
├2─ main.py
├── README.md
└3─ src
    ├4─ dataloader.py
    ├── __init__.py
    ├5─ models.py
    └6─ utils.py

1. data/: the data folder, here you have the tiny group of images overfitted
2. main.py: the main script, implements the training loop and losses
3. src/: source folder with classes
4. dataloader.py: class which implements iCoseg dataset
5. models.py: it contains SIR, CARU and MFF implementations
6. utils.py: it contains utilities for debug and tensorboard stuff

Considerations

I found the method very difficult to train, in fact i couldn't get good results on iCoseg dataset.