DGPNet - Dense Gaussian Processes for Few-Shot Segmentation

Welcome to the public repository for DGPNet. The paper is available at arxiv: https://arxiv.org/abs/2110.03674 .

How to run

Download data

1. Download and unzip PASCAL and COCO images
2. Download and unzip PASCAL and COCO annotations (we provide link here)
3. Change local_config.py to point out the images and annotations. Also change slurm_launch.sh if using slurm.
4. Download and unzip PASCAL and COCO data splits (we provide link here)
5. Make sure that the data splits are at DGPNet/data_splits

Install dependencies

The dependencies are listed in DGPNet/singularity/Dockerfile21.09

Train and test model

We typically run via slurm, using

sbatch singularity/slurm_launch.sh runfiles/dgp_5shot_pascal_resnet50.py --train --test --dataset pascal --fold 0 --add_packages_to_path

Code layout

• checkpoints - Checkpoints will be stored here at the end of training.
• data_splits - Defines the different folds.
• fss - Code is here.
• local_config.py - Used to set up paths
• logs - Used to store slurm checkpoints
• runfiles - Any experiment we run is defined in a runfile. The runfile is launched as main to start the experiment.
• singularity - We use singularity/slurm and any files related to that are stored here.
• visualization - During training and testing, our code stores some visualizations. They go here.