"Building a utopia where integrated sensing systems empower and enhance human capabilities, especially benefiting people with disabilities."
Arbitrary-modal semantic segmentation covering Depth, LiDAR, Events, Multiple Views, and RGB under 4 severe weather conditions and 5 sensor failure corner cases.
Distortion-aware Vision Transformers tailored for 360Β° panoramic cameras and omnidirectional scene understanding, featuring the 9,080-frame SynPASS benchmark.
Dynamic event-driven semantic segmentation for traffic accident scenes under extreme dynamics and motion blur, featuring the pixel-annotated DADA-seg benchmark.
End-to-end panoramic semantic mapping to indoor Bird's-Eye View (BEV) with panoramic attention mechanisms and multi-scale indoor benchmarks (Stanford-360 & Matterport-360).
Transformer-based incremental semantic mapping from egocentric observations to top-down 2D floorplans with bounding sphere attention on Matterport3D.
Hierarchical extract-and-match transformer interleaving self- and cross-attention for robust indoor (ScanNet) and outdoor (MegaDepth) local feature matching.
Efficient transformer system for transparent object segmentation and navigation assistance with Intel RealSense sensing and acoustic feedback.