Recent advancements in 3D Gaussian Splatting (3DGS) have significantly improved the mapping quality and computational efficiency of visual Simultaneous Localization and Mapping (SLAM). We propose DOGL-SLAM, a novel framework that integrates 3DGS into its core pipeline, enabling accurate camera pose tracking, object-level interaction, and high-fidelity scene reconstruction in dynamic environments. Firstly, a joint graph optimization module incorporates both dense Gaussian and sparse landmark constraints, enabling precise alignment between camera tracking and mapping. Secondly, a consistent object-level semantic fusion module embeds category labels into 3D Gaussians, grouping them via semantic consistency and optimizing distributions through cross-view losses to support scene manipulation. Finally, a hierarchical dynamic filtering pipeline is introduced, consisting of segmentation, dynamic feature exclusion, gradient masking, and visibility-aware Gaussian pruning across three parallel threads. Our system is evaluated across multiple datasets, showing a significant improvement in high-quality view synthesis for dynamic scenes. Additionally, the generated object-level semantic maps facilitate advanced downstream tasks, highlighting the potential of a robust SLAM framework.
Overview of DOGL-SLAM architecture. The system consists of five parallel threads: segmentation, tracking, local mapping, loop closing, and Gaussian
mapping. The RGB-D input is processed to extract dynamic masks and features, enabling dynamic-aware Gaussian-landmark tracking for pose estimation.
Optimized keyframes and semantics are used to refine 3D Gaussian representations through gradient masking and visibility-aware pruning, resulting in accurate,
editable, and semantically meaningful maps even in dynamic environments.
Install the docker image and clone the repo. Our Docker is based on the Dockerfile of Photo-SLAM.
docker pull wsy588/doglslam:latest
git clone https://github.com/NKU-MobFly-Robotics/DOGL-SLAM.git --recursive
# creat a container, eg:
sudo docker run -it -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=:0 --net=host -e GDK_SCALE -v /your/dataset/path:/dataset -v /your/code/path:/DOGL-SLAM -e GDK_DPI_SCALE --privileged --gpus all -e NVIDIA_DRIVER_CAPABILITIES=all --name YOURNAME wsy588/doglslam:latest /bin/bash
cd DOGL-SLAM
sh build.sh
The benchmark datasets include: TUM RGB-D, BONN. the Datasets directory is expected to have the following structure.
Datasets
├── BONN
| ├── rgbd_bonn_balloon
| ├── rgbd_bonn_crowd
| ├── rgbd_bonn_synchronous
| └── rgbd_bonn_synchronous
| ├── depth
| ├── depth.txt
| ├── groundtruth.txt
| ├── object_mask
| ├── rgb
| ├── rgb.txt
| └── dynamic_mask
└── TUM_dynamic
├── rgbd_dataset_freiburg3_walking_halfsphere
├── rgbd_dataset_freiburg3_walking_rpy
├── rgbd_dataset_freiburg3_walking_static
└── rgbd_dataset_freiburg3_walking_xyz
├── accelerometer.txt
├── depth
├── depth.txt
├── groundtruth.txt
├── object_mask
├── rgb
├── rgb.txt
└── dynamic_mask
For object_mask, instance segmentation masks can be obtained using DEVA with gaussian-grouping or directly using SAM. During segmentation, specify the category to generate a binary image dynamic_mask of the dynamic object.
# use
cd ./scripts
sh all.sh
# or one by one, eg:
cd ./scripts
../bin/tum_rgbd \
../ORB-SLAM3/Vocabulary/ORBvoc.txt \
../cfg/ORB_SLAM3/RGB-D/TUM/tum_freiburg3_walking_halfsphere.yaml \
../cfg/gaussian_mapper/RGB-D/TUM/tum_freiburg3_walking_halfsphere.yaml \
../../dataset/TUM_dynamic/rgbd_dataset_freiburg3_walking_halfsphere \
../cfg/ORB_SLAM3/RGB-D/TUM/associations/fr3_walking_halfsphere.txt \
../results/tumdy_rgbd/rgbd_dataset_freiburg3_walking_halfsphere
The evaluation is located in the DOGL-SLAM-eval directory. To run the evaluation, navigate to the DOGL-SLAM-eval directory and execute the appropriate script.
There might be minor differences between the released version and the results in the paper after developing this multi-processing version. We run all our experiments on an RTX 4090, and the performance may differ when running with a different GPU.
We sincerely thank the developers and contributors of the many open-source projects that our code is built upon.
If you found this code/work to be useful in your own research, please considering citing the following:
@ARTICLE{2026_doglslam,
title={DOGL-SLAM: Dynamic Object-Level SLAM via Joint Gaussian-Landmark Tracking},
author={Wu, Songyang and Zhang, Xuebo and Zhang, Shiyong and Yao, Runzhao and Song, Zhixing and Tong, Yunze and Yuan, Jing},
journal={IEEE Robotics and Automation Letters},
year={2026},
volume={11},
number={2},
pages={1114-1121}
}