Pytorch models and toolbox for Semantic Segmentation.

Requirements

• PyTorch >= 1.2.0
• easydict
• cv2
• skimage
• scipy
• tensorboardX
• ninja
• coco python api (if use COCO dataset)

Features

Dataset

We build many dataloaders for the popular segmentation datasets, including:

• Cityscapes
• ADE20K
• Pascal Context
• COCO

Backbone

We build 5 wellknown backbone networks, i.e., DenseNet, MobileNet, ResNet, VGG, and Xception to extract feature maps.

Module

We implement many famous and useful modules for boosting semantic segmentation performance:

• Syncronized Batch Normalization
• Atrous Spatial Pyramid Pooling (ASPP in Deeplab)
• Pyramid Pooling Modules (PPM in PSPNet)
• Gaussian Dynamic Convolution

Model

We also develop the training and testing process for some published segmentation models:

• Deeplab v3+
• PSPNet

Optimization

There are some training tricks in this repo to boost the performance:

• OHEM Loss
• Poly learning rate schedules

Visualization

In order to visualize the training and testing process. We develop a Logger module. It can log training loss and can use the Tensorboard to visualize the scalar curve and the image.

---

Performance

We evaluate the performance (mIoU) of the models in this repo on Cityscapes validation set.

Model	Papers	Ours
PSPNet	77.48	77.29
DeeplabV3+	-	79.16

---

TODO List

• Datasets

  • Pascal VOC 2012

• Modules

  • Object Context Module

• Visualization and Other Utils

  • PR-Curve
  • Segmentation Result Colorize
    • ADE20K
    • Pascal Context
    • COCO