readme.md

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

This repository contains the official implementation of the paper "DQO-MAP: Real-Time Object-Level SLAM via Dual Quadrics and Gaussians", available at here.

Abstract: 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 multi-object SLAM framework that jointly estimates the 6-DoF poses of objects and reconstructs their shapes using dual quadrics and Gaussians. Our method performs object initialization and association on the CPU and runs parallel optimization on the GPU. With object IDs as guidance, it enables fast and independent reconstruction of multiple objects in a scene. 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. Code and datasets are available at https://github.com/LiHaoy-ux/DQO-MAP.

💻1. Installation

1.1 Clone the Repository

git clone --recursive https://github.com/LiHaoy-ux/DQO-MAP
cd DQO-MAP

1.2 Python Environment

DQO-MAP was validated with Python 3.9, CUDA 11.7, and PyTorch 1.13.1; the most convenient route to install every dependency is via anaconda and pip and pipusing the following steps:

conda env create -f environment.yaml
conda activate DQO-MAP

1.3 Modified ORB-SLAM2 Python Binding

We modified ORB-SLAM2 to integrate with our ICP front-end; simply execute this script to set up pangolin, opencv, orbslam and the boost-python bindings.

bash build_orb.sh

ORB-SLAM2

Real-world sequences require ORB-SLAM2 backend optimization; install its Python bindings following the provided steps. We forked ORB_SLAM2-PythonBindings . for compatibility. Consult the original repo for build issues. Skip this step if you only test on synthetic datasets like Replica.

1.4 Test ORB-SLAM2 Python Binding

cd thirdParty/pybind/examples
python orbslam_rgbd_tum.py # please set voc_path, association_path ...
python eval_ate.py path_to_groundtruth.txt trajectory.txt --plot PLOT --verbose

Once the program completes with no errors and the trajectory matches expectations, proceed to the following stage.

📦2. Dataset Preparation

2.1 Replica

bash scripts/download_replica.sh

2.2 Room

bash scripts/download_room.sh

2.3 Dataset

|-- data
    |-- aithor
        |-- aithor1
        |-- aithor2
    |-- real
    	|-- real_world
    |-- Replica
    	|-- room0
    	|-- office0
    |-- Cube_Diorama
    	|-- room
    	

🚀3. Run

3.1 Replica

# Single Process: Recommended, More Stable
python slam.py --config ./configs/replica/office0.yaml
# Multi Process: 
python slam_mp.py --config ./configs/replica/office0.yaml

3.2 Room

# Single Process: Recommended, More Stable
python slam.py --config ./configs/Cube_Diorama/room.yaml
# Multi Process: 
python slam_mp.py --config ./configs/Cube_Diorama/room.yaml

3.3 Real

# Single Process: Recommended, More Stable
python slam.py --config ./configs/real/real.yaml
# Multi Process: 
python slam_mp.py --config ./configs/real/real.yaml

3.4 aithor

# Single Process: Recommended, More Stable
python slam.py --config ./configs/aithor/aithor1.yaml
# Multi Process: 
python slam_mp.py --config ./configs/aithor/aithor1.yaml

3.5 Set GUI

If you do not want the GUI runtime window to be displayed, set use_gui: False (default true) in the last line of config/base.yaml.

4.✈️News and TODOs

• • 📰The DQO-MAP paper published online
• • 🔥Release the rough version of the source code
• • The complete source code after open source refactoring
• • Publish the dataset online
• • A GUI page for achieving parallel and efficient mesh visualization
• • Implement an object association strategy with higher speed and accuracy

@article{li2025dqo,
  title={DQO-MAP: Dual Quadrics Multi-Object mapping with Gaussian Splatting},
  author={Li, Haoyuan and Ye, Ziqin and Hao, Yue and Lin, Weiyang and Ye, Chao},
  journal={arXiv preprint arXiv:2503.02223},
  year={2025}
}

Acknowledgments

This work extends 3DGS, leverages the ORB-SLAM2 backend from ORB-SLAM2., and adopts the Python bindings provided by ORB_SLAM2-PythonBindings. We also acknowledge the impressive contributions of RTG-SLAM and express our gratitude to all the authors for their outstanding efforts.