Grad Seminar: Advancing Semi-Supervised Domain Adaptive Semantic Segmentation Through Effective Source Integration Strategies
Presenter
Joshua Kurien, MASc candidate in Systems Design Engineering
Abstract
Semantic segmentation is a key visual recognition task with applications in medical imaging, remote sensing, and manufacturing, but developing high-performance models requires costly, dense pixel-level labels. Recent semi-supervised learning (SSL) methods built on vision foundation models offer strong segmentation performance with limited labeled data, motivating the question of how best to incorporate labeled source-domain data, like synthetic data, when available. This seminar presents a series of strategies for integrating source data into SSL pipelines, starting with simple source transfer and ensembling approaches that improve accuracy in an effective yet straightforward manner. While these methods introduce some training inefficiencies, we address them with DuCuMix, a dual-curriculum method that progressively samples source examples from easy to hard and applies a progressive mixing strategy to enhance diversity. Across synthetic-to-real and real-to-real benchmarks, this suite of methods consistently improves segmentation performance, demonstrating the value of structured source integration for label-efficient semantic segmentation.