CHARM: Collaborative Harmonization Across Arbitrary Modalities for Modality-Agnostic Semantic Segmentation

Lekang Wen, Jing Xiao*, Liang Liao, Jiajun Chen, Mi Wang (* corresponding author)

The 40th Annual AAAI Conference on Artificial Intelligence (AAAI), 2026.

TL;DR: Modality-agnostic semantic segmentation aims to achieve robust scene understanding across incomplete modality configurations. However, explicit feature alignment methods cause modal homogenization that suppresses modality-specific characteristics. CHARM enables implicit content alignment and individual enhancement through mutual perception units and dual-path optimization.

SGMA: Semantic-Guided Modality-Aware Segmentation for Remote Sensing with Incomplete Multimodal Data

Lekang Wen, Liang Liao, Jing Xiao*, Mi Wang (* corresponding author)

IEEE Transaction on Geoscience and Remote Sensing (TGRS), 2026.

TL;DR: Incomplete multimodal semantic segmentation for remote sensing integrates diverse sensors but frequently encounters missing input data. Conventional methods struggle with severe modality imbalance, massive intraclass variation, and cross-modal heterogeneity. SGMA incorporates semantic-guided fusion via classwise prototypes and modality-aware sampling to prioritize fragile modalities dynamically.

UniOrtho: From View-centric to Unified Orthographic Framework for Satellite Semantic 3D Reconstruction

Lekang Wen, Yingdong Pi*, Qiyan Luo, Jie Yang, Mi Wang (* corresponding author)

ISPRS Journal of Photogrammetry and Remote Sensing (ISPRS), 2026.

TL;DR: Satellite semantic 3D reconstruction unifies surface geometry and land classification from multi-view imagery. However, view-centric methods cause severe error accumulation and exhibit weak height-semantic coupling. UniOrtho introduces direct orthographic warping guided by explicit spatial and vertical consistency constraints.