simvisionX / weakly-segmentation-with-bounding-box

ideas from paper 'BoxSup' and 'Simple does it' and realize it on sputum smear and drone images

Example Dataset:

Sputum Smear--Makerere University, Uganda

• homepage
• paper
• code

USC Drone

• homepage
• paper

Requirements:

python3
scipy==0.19.0
numpy
pytroch>=1.1.0
torchvision>=0.3.0
PIL
opencv-python
matplotlib
lxml
[pydensecrf](https://github.com/lucasb-eyer/pydensecrf)

You can use pip to install these packages. Please add -i https://mirrors.aliyun.com/pypi/simple after package name if you are in China.

Features

Only support binary pixel classification (one object + background) now!

Model

• FCN
• UNet
• Deeplab v3+

Loss

• BCE
• Focal Loss
• Dice Loss
• Lovase Loss

Pseudo segmentation label generation

• all bounding box
• inner area of bounding box
• grabcut

Dense CRF for post-process

Training

• sgd
• adam
• update label for iteration training
• mixup

Further work

• support multi-classes
• more fast and simple mIoU calculation
• more useful model
• more appropriate optimizer

Example of this project: