Research map / Label-efficient and robust learning / Domain adaptation and generalization

MaskTwins: Dual-form Complementary Masking for Domain-Adaptive Image Segmentation

ICML 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Adaptation for semantic segmentation · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MaskTwins, a UDA segmentation framework that enforces consistency between predictions on complementary (dual-form) masked target images, with a theoretical argument for complementary over random masks.

Problem

Masked image consistency in unsupervised domain adaptation (UDA) for segmentation lacks theoretical grounding, and its potential for domain-invariant feature learning is underexploited, especially within a single modality.

Approach

Reframes masked reconstruction as sparse signal reconstruction and argues that complementary mask pairs (D, 1-D) preserve more information and give tighter bounds than random masks; uses these pairs in an EMA-teacher self-training loop with a complementary-mask consistency loss, with no extra learnable parameters.

Key results

• On SYNTHIA->Cityscapes, MaskTwins reaches 76.7 mIoU versus 74.0 for MIC (the prior best in the table).
• On VNC III->Lucchi V2L1 and V2L2, IoU is 75.0 and 78.6, improving on the state of the art by about 3.2 points each.
• On WASPSYN, the F1-score is 0.5913 versus 0.5711 for MIC.

Measured effects reported in the paper

Surprising findings

• The paper's own proof gives a zero information-preservation value for complementary masks and a positive mean of 1/4 for random masks. That is the opposite of the claimed advantage, which the paper does not reconcile.

Limitations

• Gains on small objects are smaller or negative, which the authors attribute to complementary masked regions confusing the model.
• The theoretical analysis rests on strong assumptions (sparse signal model, RIP, Lipschitz f), and the proofs contain gaps, for example a rate step in the Rademacher bound with no justification.
• Evaluated only on segmentation tasks; the authors note that classification and video are untested.

Experimental setup

Datasets: SYNTHIA Cityscapes VNC III Lucchi MitoEM WASPSYN

Benchmarks: SYNTHIA->Cityscapes VNC III->Lucchi MitoEM-R->MitoEM-H WASPSYN

Metrics: mIoU IoU F1 MCC mAP

Baselines: DAFormer CAMix HRDA MIC UALR DAMT-Net DA-VSN DA-ISC CAFA SSNS-Net AdaSyn

Builds on: MIC DAFormer HRDA AdaIN

Closest papers by meaning

• Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.87)
• Reconstruction Target Matters in Masked Image Modeling for Cross-Domain Few-Shot Learning (AAAI 2025 · similarity 0.86)
• Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (AAAI 2025 · similarity 0.85)
• MaskViM: Domain Generalized Semantic Segmentation with State Space Models (AAAI 2025 · similarity 0.85)
• Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.84)
• ChA-MAEViT: Unifying Channel-Aware Masked Autoencoders and Multi-Channel Vision Transformers for Improved Cross-Channel Learning (NeurIPS 2025 · similarity 0.84)