Tensorflow 2.1 Segmentation pipeline implementation from scratch

Process with Cityscape datasets

Cityscape data with HRNet

cityscape dataset (https://www.cityscapes-dataset.com/)
HRNet (https://arxiv.org/pdf/1908.07919.pdf)

HRNetV2W48 model
mIOU 70.31% (without pre-trained backbone)

| Input image | Ground truth | Segmentation result |

Process with Inria datasets

Inria dataset with *Modified HRNet

inria dataset (https://project.inria.fr/aerialimagelabeling/leaderboard/)
HRNet (https://arxiv.org/pdf/1908.07919.pdf)
*Modify HRNet's layers for satellite image processing

| Input image | Ground truth | Segmentation result |

Foreground IOU : 75.63% (without post-processing)