Object-Aware SLAM Based on Efficient Quadric Initialization and Joint Data Association, Zhenzhong Cao, Yunzhou Zhang, Rui Tian, Rong Ma, Xinggang Hu, Sonya Coleman, Dermot Kerr | Project

• We propose an efficient quadric initialization (EQI) method based on object detection and surfel construction which initializes quadrics using fewer frames with small viewing angles.
• We propose a robust object-level joint data association (JDA) method combining multi-dimensional information and statistic distributions.
• We propose a multi-constraint optimization factor graph for quadrics optimization and joint bundle adjustment.
• We implement a complete visual semantic SLAM system, aiming to build a novel object-oriented and semanticallyenhanced map for indoor robot interaction.

SemanticTopoLoop: Semantic Loop Closure With 3D Topological Graph Based on Quadric-Level Object Map, Zhenzhong Cao | Project

• MLV-ODA method is introduced to reduce the time and space complexity of data association, indirectly promoting the accuracy and completeness of object construction in the scene.
• QLT-SLC method is presented to improve the precision and recall rate of loop closure, as well as enhance the system’s localization accuracy.
• The proposed MLV-ODA method and QLT-SLC method are embed into the Object-Aware SLAM system, which jointly maintain the PPO-MD.
• Qualitative experiments, quantitative experiments, and ablation studies are designed to demonstrate the effectiveness and robustness of the proposed MLV-ODA and QLT-SLC method.

RGBDS-SLAM: A RGB-D Semantic Dense SLAM Based on 3D Multi Level Pyramid Gaussian Splatting, Zhenzhong Cao | Project

• We introduce a 3D Multi-Level Pyramid Gaussian Splatting (MLP-GS) method, which extracts multi-level image pyramids for Gaussian splatting training to restore scene details and ensure reconstruction consistency.
• We design a Tightly Coupled Multi-Features Reconstruction Optimization (TCMF-RO) mechanism to mutually boost RGB, depth, and semantic reconstruction accuracy during rendering optimization.
• We build a full RGB-D semantic dense SLAM system extended from ORB-SLAM3, supporting real-time high-fidelity joint reconstruction of color, depth and semantics.
• Experiments on Replica and ScanNet datasets surpass SOTA methods, with 11.13% PSNR and 68.57% LPIPS improvement; ablation experiments verify the effectiveness of two core modules.

LLF-GS SLAM: Real-Time 3D Gaussian Semantic SLAM via Compact Label-Language Features, Zhenzhong Cao | Project

• We propose a label-language dual-branch semantic Gaussian representation: global consistent explicit labels act as object segmentation anchors, and each label stores a compact multi-view language feature bank for open-vocabulary query and editing.
• A global label consistency mechanism aligns per-frame SAM segmentation with rendered semantic maps to eliminate cross-view label fragmentation.
• We compress 512-dim CLIP features into 16-dim compact embeddings via offline autoencoder, and design quality-aware update rules for feature banks to save memory.
• The system runs in real time on low-memory GPUs, supporting open-vocabulary 3D object retrieval and interactive object editing; extensive experiments validate its superiority on Replica and TUM RGB-D.