TACS

Guide: Prof. Balamurugan Palaniappan

This work was done as part of the paper review in the course IE 643 Deep Learning - Theory and Practice. I reviewed the literature for the problem of adapting semantic segmentation models to a domain of labels, going through the works ADVENT, Open Source Adaptation, and TACS.

Replicated the results of TACS using the open-source code, after a detailed survey of the paper, and implemented a modification using a post-training ensemble-based method to improve the mean IoU percentage by 2-8% over different taxonomies (up to 25-30% increase in mean IoU percentage for the worst performing labels).