DrivingEditor: 4D Composite Gaussian Splatting for Reconstruction and Edition of Dynamic Autonomous Driving Scenes.

Area of Science:

• Computer Vision
• Robotics
• 3D Scene Reconstruction

Background:

• Autonomous driving necessitates accurate 3D reconstruction of large-scale, unbounded scenes.
• Current 3D reconstruction methods for autonomous driving often lack scene editing capabilities.
• Existing scene editing methods rely heavily on manual 3D bounding box annotations, limiting scalability.

Purpose of the Study:

• To develop a novel 3D scene reconstruction method for autonomous driving that supports scene editing.
• To overcome the limitations of scalability and manual annotation in current scene editing techniques.
• To enhance the reconstruction quality of both dynamic and static elements in autonomous driving scenarios.

Main Methods:

• Introduced DrivingEditor, a new Gaussian representation for 3D scene modeling.
• Decoupled scene modeling into separate branches for dynamic foreground objects and static backgrounds.
• Utilized a framework for decoupled scenario modeling to enable individual object manipulation.

Main Results:

• Achieved accurate editing of dynamic targets, including object removal and addition, without 3D bounding boxes.
• Improved the reconstruction quality of autonomous driving scenes, particularly for dynamic foreground objects.
• Demonstrated robust performance on Waymo Open Dataset and KITTI benchmarks for both dynamic and static scenes.
• Showcased strong performance and robustness in unstructured large-scale scenarios.

Conclusions:

• DrivingEditor offers a scalable and effective solution for 3D scene reconstruction and editing in autonomous driving.
• The decoupled modeling approach enhances both reconstruction accuracy and editing flexibility.
• The method shows significant potential for applications requiring dynamic scene manipulation and high-fidelity reconstruction.