SegBench: A simple toolbox for Implementing Binary Segmentation

Introduction

This project aims to build up a benchmark for fair and solid comparisons with the state-of-the-art segmentation methods. Motivated by sssegmentation, we hope to excel at the Binary Segmentation which is prevailing in minority tasks but also matters, e.g., Shadow Detection, Camouflage Object Detection, Saliency Object Detection, and most of Medical Image Segmentation.

In a sum, we:

• Present read and deployment-friendly codes for Training/Testing
• Present a wide range of cutting-edge methods comparisons with several datasets
• Provide full experiment details including the logs, tensorboard（too large, forget it 😅）, and ckpt

Give me some time to construct this project :)

Environments

The codes mainly rely on the HuggingFace🤗, including the Distributed Training (Accelerate), Model Architecture (Transformers).

My stable version for these packages are:

pip install accelerate==0.20.3
pip install transformers

Pytorch is suggested to update with version 1.12.1 or higher. I will try more efficient implementations with Flash Attention , Torch Compile and etc. in the future.

Support Datasets

Shadow Detection

• BIGSHA(coming soon): High Resolution Shadow Detection
• SBU (ECCV16): Image Shadow Detection
• CUHK-Shadow (TIP21): Image Shadow Detection, larger one

Camouflaged object detection

• COD10K (CVPR20): Image Camouflaged object detection

Transparent Segmentation

• Trans10k (ECCV20) : Transparent objects segmentation:

Medical Image Segmentation

• ISIC [RGB image] : Skin Lesion Segmentation,
• BUSI [ultrasound] : Breast Cancer Ultrasound Image Segmentation
• BCN[pathology]: Breast Cancer Mitosis Nuclei Segmentation
• MHSI[pathology] : Melanoma Segmentation
• GlaS [pathology] : Intestinal Glandular Structures Segmentation

And So On

Detailed Documents

Please check the corresponding doc to find the model card, implement configs, and visualization

• Interest on Natural Scenario ? 👉️👉️👉️ Check Natural Scenario Documents (TODO)
• Interest on Medical Images ? 👉️👉️👉️ Check Medical Documents (TODO)

Future works

So many blanks need to fill :)

Keep Patience :)

Drop emails to haipengzhou856@gmail.com or directly post the issues here if you have any questions.

Acknowledgement

Thanks for all the open-source and code-sharing contributors.

Please considering to cite those datasets and the reproduced methods.

This codebase is built on HuggingFace🤗.

License

Currently, the project is under CC BY-NC 2.0, Any kinds of modification is welcome.

But Forbidden Commercial Usage.

All Copyright © Rydeen, Haipeng ZHOU