Reproduce “Instance-Aware Test-Time Segmentation for Continual Domain Shifts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seunghwan Lee, Inyoung Jung, Hojoon Lee, Eunil Park, Sungeun Hong
CoTICA, a framework for semantic segmentation, continually adapts to evolving domain shifts during test time by dynamically adjusting pseudo-label thresholds per instance and class. The method also prioritizes learning for challenging classes, achieving a 6.2% mIoU gain over the pre-trained source model on the ACDC dataset.
Key insights
- Dynamic, fine-grained, instance- and class-specific pseudo-label thresholds (ICAT) are crucial for generating reliable supervision and achieving robust semantic segmentation under continual domain shifts.
- Strategically weighting the loss based on temporal, class-wise difficulty (ICWL) helps mitigate error accumulation and improves adaptation stability for persistently challenging categories.
- Combining adaptive thresholding with a weighted loss, as in CoTICA, allows models to better manage both the quality of self-supervision and the focus of learning in dynamic environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.