sarthak-chakraborty/Image-Segmentation-UNET-DeepLab

Image Segmentation on 3 popular datasets using Unet and DeepLab model

Image Segmentation

Datasets Used

• Cityscapes
• PascalVOC 2012
• CamVid

Requirements

pip install -r requirements.txt

Dataset Preprocessing

Follow steps in dataset_preprocess and accumulate the data in a particular directory structure. To test with other datasets, you will need to follow the same directory structure as mentioned in dataset_preprocess.

Models Used

1. Unet (tested with all datasets). Follow these materials to get a better understanding of the Unet architecture:

   • U-Net: Convolutional Networks for Biomedical Image Segmentation
   • Understanding Semantic Segmentation with UNET

2. DeepLabv3 model (tested only for PascalVOC 2012 dataset). To know about DeepLabv3 architecture, follow these materials:

   • Rethinking Atrous Convolution for Semantic Image Segmentation
   • DeepLabv3: Semantic Image Segmentation

Run

Run the notebook segmentation.ipynb in steps to get the desired results. To run with DeepLab, run the notebook DeepLab/segmentation.ipynb.

References

1. Multiclass Semantic Segmentation Camvid
2. Keras DeepLab V3.1+