Semi - Supervised Learning for Semantic Segmentation using Limited Point Annotation

In this Repo I overcome the challenge of limited point annotation by:

• Using pandas and OpenCV to prepare the satellite images and masks from the DeepGlobe dataset, ensuring they are appropriately formatted for the model.

• Setting up the DeepLabV3Plus model with ResNet-50 as the backbone, leveraging transfer learning by initializing with pre-trained ImageNet weights using the segmentation_models_pytorch library.

• Using NumPy to simulate random point labels on segmentation masks to handle partially labeled data effectively.

• Varying sampling ratio and image resolution to study their effects on the model's Intersection over Union (IoU) performance.