DQO-MAP: Real-Time Object-Level SLAM with Dual Quadrics and Gaussians

While 3D Gaussian Splatting has recently demonstrated impressive results in scene-level reconstruction, it lacks the ability to reconstruct and track individual objects with accurate poses. This limitation hinders its application in downstream tasks such as augmented reality and embodied AI, where object-level understanding is essential. To address this gap, we propose DQO-MAP, a real-time online object-level SLAM system that jointly estimates the 6-DoF poses of objects and reconstructs their shapes using dual quadrics and Gaussians. With a robust 3D-2D object association framework for globally coherent mapping and an incremental 3D Gaussian update strategy with an efficient object-level loss and multi-view visibility pruning, our system achieves improved geometric accuracy and real-time performance. Experimental results on Cube-Diorama, Replica, and self-collected datasets demonstrate that DQO-MAP achieves higher reconstruction accuracy, completeness, and pose precision while running faster and with lower memory than prior methods.