/Seg-UNet

Ensemble architecture of SegNet and UNet for Semantic Segmentation with keras

Primary LanguagePython

Seg-UNet(SegNet + UNet)

SegUNet is a model of semantic segmentation based on SegNet and UNet(these model are based on Fully Convolutional Network). Architecture dedicated to restoring pixel position information. This architecture is good at fine edge restoration etc.

This repository contains the implementation of learning and testing in keras and tensorflow.

Architecture

  • SegNet

    SegNet image

    • indoces pooling

      indicespoolingimage

  • UNet

    • skip connection

      UNet image

This architecture is encoder-decoder model(29 conv2D layers).

  • Skip connection(UNet) and indeces pooling(SegNet) are incorporated to propagate the spatial information of the image.

Usage

train

  • Segmentation involveing multiple categories

    python train.py --options

  • Segmentation of mask image

    python train_mask.py --options

    • options

      • image dir
      • mask image dir
      • batchsize, nb_epochs, epoch_per_steps, input_configs
      • class weights
      • device num

DEMO

  • dataset

    • LIP(Look into person)

      demo1

Author

ykamikawa