Learning to Model Pixel-Embedded Affinity for Homogeneous Instance Segmentation

Accepted by AAAI-2022

Wei Huang, Shiyu Deng, Chang Chen, Xueyang Fu, Zhiwei Xiong(*)

*Corresponding Author

University of Science and Technology of China (USTC)

Introduction

This repository is the official implementation of the paper, "Learning to Model Pixel-Embedded Affinity for Homogeneous Instance Segmentation", where more implementation details are presented.

Installation

This code was tested with Pytorch 1.0.1 (later versions may work), CUDA 9.0, Python 3.7.4 and Ubuntu 16.04. It is worth mentioning that, besides some commonly used image processing packages, you also need to install some special post-processing packages for neuron segmentation, such as waterz and elf.

If you have a Docker environment, we strongly recommend you to pull our image as follows,

docker pull registry.cn-hangzhou.aliyuncs.com/renwu527/auto-emseg:v5.4

or

docker pull renwu527/auto-emseg:v5.4

Dataset

Datasets	Training set	Validation set	Test set	Download (Processed)
CVPPP (A1)	530x500x108	530x500x20	530x500x33	BaiduYun (Access code: weih) or GoogleDrive
BBBC039V1	520x696x100	520x696x50	520x696x50	BaiduYun (Access code: weih) or GoogleDrive
AC3/AC4	1024x1024x80	1024x1024x20	1024x1024x100	BaiduYun (Access code: weih) or GoogleDrive

Download and unzip them in corresponding folders in './data'.

CVPPP

cd ./scripts_cvppp

0. Install metrics for CVPPP

cd ./lib/evaluate
python setup.py build_ext --inplace
cd ../..

1. Training

python main.py -c=cvppp

2. Inference

python inference.py -c=cvppp -mn=cvppp -id=cvppp-validation -m=validation

BBBC039V1

cd ./scripts_bbbc039v1

1. Training

python main.py -c=bbbc039v1

2. Inference

python inference.py -c=bbbc039v1 -mn=bbbc039v1 -id=bbbc039v1-test -m=test

AC3/AC4

cd ./scripts_ac3ac4
pip install waterz

1. Training

python main.py -c=ac3ac4

2. Inference

python inference.py -c=ac3ac4 -mn=ac3ac4 -id=ac3ac4-test -m=ac3

Notice for CVPPP

During the training phase, we use the SBD metric (./scripts_cvppp/utils/evaluate.py/SymmetricBestDice) to select models.

In our paper, we provide the SBD score and the absDiffFG score returned by the website as follows.

ResNet-50	ResNet-101	ResUNet

However, we have noticed that the website made adjustments to the SBD metric at the end of 2021.

Therefore, we submit our previous results again (05/02/2022), and obtain the new return scores as follows.

ResNet-50	ResNet-101	ResUNet

We can find that the new SBD scores are significantly lower than the previous scores.

Contact

If you have any problem with the released code, please do not hesitate to contact me by email (weih527@mail.ustc.edu.cn).