• computer vision
• deep learning
• remote sensing
• semantic segmentation
• weakly supervised learning

In this thesis, we present a study on weakly supervised semantic segmentation in
aerial imagery, focusing on the potential of the "Segment Anything Model" (SAM),
a novel foundation model, to facilitate the training of a semantic segmentation model
without requiring complete pixel-level annotations.
Our approach leverages bounding boxes as weak annotations for training the model.
The performance of this method is compared with that of the same model trained
using original annotations.
Furthermore, we explore SAM’s capability to enhance the results of a semantic
segmenter. Given SAM’s ability to provide highly accurate contour masks, we post-
process the outputs of the weakly supervised trained model to get performance
improvement.