Reproduce “MC-PanDA: Mask Confidence for Panoptic Domain Adaptation” 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
Ivan Martinović, Josip Šarić, Siniša Šegvić
Researchers from the University of Zagreb developed MC-PanDA, a framework that enhances unsupervised domain adaptation for panoptic segmentation by integrating mask-wide and pixel-level confidence from mask transformers into the self-training process. The approach achieved a 6.2 PQ point improvement over previous methods on the Synthia→Cityscapes benchmark, establishing new state-of-the-art performance.
Key insights
- Mask transformers inherently possess the capability to estimate prediction uncertainty at a granular, fine-grained level (mask-wide and pixel-level).
- Leveraging this fine-grained confidence is crucial for mitigating noise amplification in pseudo-labels and enabling more robust and stable self-training in unsupervised domain adaptation.
- Mask-wide Loss Scaling (MLS) and Confidence-based Point Filtering (CBPF) offer complementary mechanisms that together guide the student model to learn cautiously from high-confidence teacher predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.