Bohan Ren

CubifyGS is an object-centric mapping framework for lifelong scene maintenance in dynamic environments with frequent rigid object rearrangement. Instead of relying on slow primitive-level re-optimization that often leaves persistent ghosting artifacts, it updates 3D Gaussian assets through spatio-temporal dynamics perception, global asset retrieval, rigid transformation, and explicit pruning. An event-triggered adaptive optimization strategy then refines only the affected regions, improving both artifact suppression and long-term maintenance efficiency on a newly constructed high-fidelity dynamic benchmark.

OpenGS-Fusion is an open-vocabulary dense mapping framework that couples 3D Gaussian Splatting with a TSDF representation for on-the-fly, lossless semantic fusion. It further introduces MLLM-Assisted Adaptive Thresholding to refine 3D object segmentation under open-ended language queries, substantially improving object-level understanding and delivering stronger scene reconstruction and language-guided interaction performance than prior methods.

This work presents an automated framework for 3D Gaussian Splatting sub-map registration and fusion without manual reference-map selection. By combining cross-scene geometric skeleton extraction with ellipsoid-aware convolution and a multi-factor Gaussian fusion strategy, it improves both alignment accuracy and post-fusion rendering quality, delivering more reliable 3D scene representation for robotic perception and autonomous navigation.

This work targets RGB-thermal semantic segmentation in snowy urban scenes by addressing both the lack of training data and practical deployment constraints. It introduces the SUS dataset and a knowledge-distillation framework built around MCNet, where prior thermal knowledge from an RGB-T teacher is transferred to an efficient RGB student, leading to stronger segmentation performance than existing methods on both SUS and MFNet.