Region Proportion Regularized Inference (RePRI) for Few-Shot Segmentation

In this repo, we provide the code for our paper : "Few-Shot Segmentation Without Meta-Learning: A Good Transductive Inference Is All You Need?", available at https://arxiv.org/abs/2012.06166:

Getting Started

Minimum requirements

1. Software :

• torch==1.7.0
• numpy==1.18.4
• cv2==4.2.0
• pyyaml==5.3.1

For both training and testing, metrics monitoring is done through visdom_logger (https://github.com/luizgh/visdom_logger). To install this package with pip, use the following command:

pip install git+https://github.com/luizgh/visdom_logger.git

2. Hardware : A 11 GB+ CUDA-enabled GPU

Download data

All pre-processed from Google Drive

We provide the versions of Pascal-VOC 2012 and MS-COCO 2017 used in this work at https://drive.google.com/file/d/1Lj-oBzBNUsAqA9y65BDrSQxirV8S15Rk/view?usp=sharing. You can download the full .zip and directly extract it at the root of this repo.

If the previous download failed

Here is the structure of the data folder for you to reproduce:

data
├── coco
│   ├── annotations
│   ├── train
│   ├── train2014
│   ├── val
│   └── val2014
└── pascal
|    ├── JPEGImages
|    └── SegmentationClassAug

Pascal : The JPEG images can be found in the PascalVOC 2012 toolkit to be downloaded at PascalVOC2012 and SegmentationClassAug (pre-processed ground-truth masks).

Coco : Coco 2014 train, validation images and annotations can be downloaded at Coco. Once this is done, you will have to generate the subfolders coco/train and coco/val (ground truth masks). Both folders can be generated by executing the python script data/coco/create_masks.py (note that the script uses the package pycocotools that can be found at https://github.com/cocodataset/cocoapi/tree/master/PythonAPI/pycocotools):

python

cd data/coco
python create_masks.py

About the train/val splits

The train/val splits are directly provided in lists/. How they were obtained is explained at https://github.com/Jia-Research-Lab/PFENet

Download pre-trained models

Pre-trained backbones

First, you will need to download the ImageNet pre-trained backbones at https://drive.google.com/drive/folders/1Hrz1wOxOZm4nIIS7UMJeL79AQrdvpj6v and put them under initmodel/. These will be used if you decide to train your models from scratch.

Pre-trained models

We directly provide the full pre-trained models at https://drive.google.com/file/d/1iuMAo5cJ27oBdyDkUI0JyGIEH60Ln2zm/view?usp=sharing. You can download them and directly extract them at the root of this repo. This includes Resnet50 and Resnet101 backbones on Pascal-5i, and Resnet50 on Coco-20i.

Overview of the repo

Data are located in data/. All the code is provided in src/. Default configuration files can be found in config_files/. Training and testing scripts are located in scripts/. Lists/ contains the train/validation splits for each dataset.

Training (optional)

If you want to use the pre-trained models, this step is optional. Otherwise, you can train your own models from scratch with the scripts/train.sh script, as follows.

bash scripts/train.sh {data} {fold} {[gpu_ids]} {layers}

For instance, if you want to train a Resnet50-based model on the fold-0 of Pascal-5i on GPU 1, use:

bash scripts/train.sh pascal 0 [1] 50

Note that this code supports distributed training. If you want to train on multiple GPUs, you may simply replace [1] in the previous examples with the list of gpus_id you want to use.

Testing

To test your models, use the scripts/test.sh script, the general synthax is:

bash scripts/test.sh {data} {shot} {[gpu_ids]} {layers}

This script will test successively on all folds of the current dataset. Below are presented specific commands for several experiments.

Pascal-5i

Results :

Command:

bash scripts/test.sh pascal 1 [0] 50  # 1-shot
bash scripts/test.sh pascal 5 [0] 50  # 5-shot

Coco-20i

Results :

Command :

bash scripts/test.sh coco 1 [0] 50  # 1-shot
bash scripts/test.sh coco 5 [0] 50  # 5-shot

Coco-20i -> Pascal-VOC

The folds used for cross-domain experiments are presented in the image below:

Results :

Command :

bash scripts/test.sh coco2pascal 1 [0] 50  # 1-shot
bash scripts/test.sh coco2pascal 5 [0] 50  # 5-shot

Monitoring metrics

For both training and testing, you can monitor metrics using visdom_logger (https://github.com/luizgh/visdom_logger). To install this package, simply clone the repo and install it with pip:

git clone https://github.com/luizgh/visdom_logger.git
pip install -e visdom_logger

Then, you need to start a visdom server with:

python -m visdom.server -port 8098

Finally, add the line visdom_port 8098 in the options in scripts/train.sh or scripts/test.sh, and metrics will be displayed at this port. You can monitor them through your navigator.

Contact

For further questions or details, please post an issue or directly reach out to Malik Boudiaf ([email protected])

Acknowledgments

We gratefully thank the authors of https://github.com/Jia-Research-Lab/PFENet, as well as https://github.com/hszhao/semseg from which some parts of our code are inspired.