/DeepLab_v3_plus

This is an ongoing re-implementation of DeepLab_v3_plus on pytorch which is trained on VOC2012 and use ResNet101 for backbone.

Primary LanguagePythonMIT LicenseMIT

DeepLab_V3_plus : a model about semantic segmentation

This is a simple pytorch re-implementation of Google Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation.

Introduction:

This work still need to be updated. The features are summarized blow:

  • Use ResNet101 as base Network. Xception will be updated soon.
  • Use only VOC2012 for base dataset. Other dataset will be updated soon.

We have finished:

  • Version avaliable for VOC2012.

  • You can check your dataloader error in 'path/to/workspace/check/check_dataloader/img'. You will see three part:

    • 1.original image which we load directly from image path.
    • 2.restore image from torch-tensor(transformed from numpy) back to numpy ndarray.
    • 3.mask loaded by dataloader. Following images show original image, augumentation image and mask target from left to right
  • Network architecture.

  • Evaluation mIOU on PASCAL VOC2012 valset every epoches.

  • Pretrained model on PASCAL VOC2012 BaiduYun Link, which is trained on 2 Tesla P100 for 100 epoches with config shown in code.

Usage:

  • Download dataset and unzip
ln -s VOCdevkit path/to/deeplab_v3_plus/dataset
  • Pretrained model is avaliable BaiduYun Link
  • Finally, run the model.
  • Check you GPU resources and modify your run.sh.
sh run.sh

Future:

  • Xception as Network Baseline.
  • Pretrained model on COCO, JFT.
  • Depthwise separable convolution.
  • Support dataset for Cityscapes.
  • Visualization for test result and gt.