Project
STAS (Medical Image Segmentation Challenge)
Repository for our award-winning solution in a medical image segmentation competition, including training, inference, and experiment configurations for reproducible results.
Research Artifacts (Pipelines)
Below are simplified pipeline overviews for selected papers. Each figure highlights the main idea and the system-level flow.
A robustness-oriented deepfake detector that learns with multi-task supervision. The pipeline encourages the model to capture forgery-consistent artifacts by jointly optimizing complementary objectives (e.g., forgery cues and auxiliary predictions), improving generalization across manipulation types and datasets.
A practical segmentation framework for limited labels: cross-domain augmentation expands appearance diversity, while model ensembling stabilizes predictions. The pipeline is designed for real-world clinical variability and improves robustness under limited supervision.
A training-free / few-shot industrial anomaly detection pipeline that produces patch-level anomaly maps for verification. The core idea is to unify visual prompting signals and patch-level comparisons to localize anomalies, enabling reliable inspection with minimal data and strong interpretability.