PointScatter based on MMSegmentation

This is an implementation of the paper PointScatter: Point Set Representation for Tubular Structure Extraction (ECCV2022 Oral).

Installation

1.First install MMSegmentation following mmseg_github.

2.To install pointscatter, run python setup.py develop

Dataset Preparation

The location of the datasets should be the same as the path indicated in the corresponding file in "./configs/_base_/datasets/".

For DRIVE and STARE, refer to mmsegmentation for preparation process.

Massachusetts Roads can be downloaded from massroad_kaggle. We select a subset of 694 images to train, which is indicated in "./configs/_base_/datasets/massroads_split/", and test the model on the 13 official test images.

DeepGlobe can be downloaded from deepglobe_challenge. We split it by "./configs/_base_/datasets/deepglobe_split/".

You may need to prepare the two road datasets according to the following structure after downloading them.

mass_roads_choose
│
└───training
│   └───input
│   └───output
│   └───output_cline
└───testing
│   └───...

Train and Test

Train pointscatter with a single GPU:

python train.py configs/segmentors/3_pointscatter.py --dataset drive --backbone unet --work-dir ../output/drive_log/unet_3_pointscatter

Train with multiple GPUs:

python -m torch.distributed.run --nproc_per_node=4 --master_port=8888 train.py configs/segmentors/3_pointscatter.py --dataset drive --backbone unet --work-dir ../output/drive_log/unet_3_pointscatter --launcher pytorch

Test a trained model:

python test.py configs/segmentors/3_pointscatter.py  ../output/drive_log/unet_3_pointscatter/latest.pth --dataset drive --backbone unet --work-dir ../output/drive_log/unet_3_pointscatter --eval mDice

Pretrained Models can be found:
Google Drive
Baidu Cloud

Customize models

You can customize your own model by:

1. Add a new backbone in "./pointscatter/modeling/backbones" or register a new dataset in "./pointscatter/datasets".
2. Configure the new model components in "./configs/".
3. Train the model.