Image Segmentation | Class Activation Maps | ViTs
This project was completed as part of the Deep Neural Networks (DNN) module in SMU. The goal was to segment pixels within an image into different classes based off Class Activation Maps (features identified as belonging to a certain class with high probability).
The training data for this was images without any labelled pixels. To be specific, this branch of image segmentation falls in the weakly supervised category - meaning that only the image classification labels are given but not the pixel level classification.
The full report containing the methods we used and the different models used to identify features and subsequently classify pixels is appended below.
