ML_ScratchPad

Semantic Segmentation of Images using Popular FCN based Architectures

Following Models are implemented

• Attention U-Net
• Deep Dilation Net

The Potsdam Dataset (RGBIR + DSM & Labels) is also uploaded and divided into training, validation and testing (no labels available).

The images and corresponding labels are indexed in the form of hash table stored in a json file. Custom loss functions are also implemented accordingly.

Requirements

• tensorflow >= 2.1
• keras
• numpy
• rasterio

NOTE: Please update the file paths in the vrt files to POSIX path or Windows path according to your platform, i.e. you need to replace \ with / or vice versa in the path strings.