/pkuseg

Benchmarks for Semantic Segmentation.

Primary LanguagePythonApache License 2.0Apache-2.0

PKUSeg

Introduction

PKUSeg is an open source semantic segmentation toolbox based on PyTorch, which is maintained by EECS of Peking University. Maintainers are all from Key Laboratory of Machine Perception (MOE).

Major features

  • Modular design and easy to use and deploy
    We develop this tool for easier experiments and deployment．
  • All kinds of models for semantic segmentation
    We implement many state-of-the-art models in research papers. We not only release codes, but also training checkpoints.
  • State-of-the-art results on multiple datasets
    We achieve the state-of-the-art results on multiple datasets including Pascal VOC, Cityscapes, Pascal Context and ADE20K．

Implemented Papers

  • PSPNet: Pyramid Scene Parsing Network CVPR2017
  • DeepLabV3: Rethinking Atrous Convolution for Semantic Image Segmentation CVPR2017
  • DeepLabV3+: Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation ECCV2018
  • DenseASPP: DenseASPP for Semantic Segmentation in Street Scenes CVPR2018
  • DANet: Dual Attention Network for Scene Segmentation CVPR2019
  • EMANet: Expectation-Maximization Attention Networks for Semantic Segmentation ICCV2019

Performances with PKUSeg

All the performances showed below fully reimplemented the papers' results.

PASCAL VOC

  • Single Scale Whole Image Test: Base LR 0.01, Crop Size 513
Model Backbone Train Test mIOU BS Iters Scripts
PSPNet 3x3-Res101 train val 78.75 16 3W PSPNet
DeepLabV3 3x3-Res101 train val 78.95 16 3W DeepLabV3
EMANet 3x3-Res101 train val 79.79 16 3W EMANet

Cityscapes

  • Single Scale Whole Image Test: Base LR 0.01, Crop Size 769
Model Backbone Train Test mIOU BS Iters Scripts
PSPNet 3x3-Res101 train val 78.20 8 4W PSPNet
DeepLabV3 3x3-Res101 train val 79.13 8 4W DeepLabV3

ADE20K

  • Single Scale Whole Image Test: Base LR 0.02, Crop Size 520
Model Backbone Train Test mIOU PixelACC BS Iters Scripts
PSPNet 3x3-Res50 train val 41.52 80.09 16 15W PSPNet
DeepLabv3 3x3-Res50 train val 42.16 80.36 16 15W DeepLabV3
PSPNet 3x3-Res101 train val 43.60 81.30 16 15W PSPNet
DeepLabv3 3x3-Res101 train val 44.13 81.42 16 15W DeepLabV3

License

This project is released under the Apache 2.0 license.