Multi-Spectral Learning

Fusion and cross-modal learning across RGB, infrared, and SAR imaging modalities for robust perception.

Multi-spectral learning exploits complementary information across imaging modalities — RGB, infrared (IR), and Synthetic Aperture Radar (SAR) — to build perception systems that are robust under challenging conditions such as low light, fog, and adverse weather. Our work addresses cross-modal fusion, missing-modality robustness, spectral knowledge distillation, and semi-supervised image translation, targeting applications in robotic perception, aerial surveillance, and autonomous systems.

OGP-Net: Optical Guidance Meets Pixel-Level Contrastive Distillation for Robust Multi-Modal Segmentation (AAAI 2025):

OGP-Net combines optical guidance with pixel-level contrastive distillation to achieve robust multi-modal segmentation even when one or more input modalities are missing at test time.

SKD-Net: Spectral-Based Knowledge Distillation in Low-Light Thermal Imagery (ICRA 2024):

SKD-Net uses spectral-domain knowledge distillation to transfer perceptual knowledge from RGB to thermal imagery, significantly improving robotic perception in low-light environments.

SSL-RGB2IR: Semi-Supervised RGB-to-IR Image Translation (IROS 2024):

SSL-RGB2IR translates RGB images to IR using a semi-supervised framework, enabling training data augmentation for both semantic segmentation and object detection without requiring paired IR annotations.