DroneSplat+: Semantics-Enhanced 3D Gaussian Splatting for Robust 3D Reconstruction from In-the-Wild Drone Imagery

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Drones offer agile aerial perspectives for scene reconstruction.
• Radiance field methods enable photorealistic rendering but struggle with limited views and dynamic elements in drone imagery.
• Existing challenges in geometric recovery and multi-view consistency are coupled in real-world drone data.

Purpose of the Study:

• To introduce DroneSplat+, a framework addressing coupled challenges in 3D reconstruction from drone imagery.
• To improve geometric accuracy under limited-view constraints.
• To effectively handle dynamic objects and ensure scene consistency.

Main Methods:

• DroneSplat+ integrates multi-view stereo (MVS) with semantic segmentation priors to guide Gaussian optimization.
• A dual-elimination strategy, including adaptive masking and label-based localization, identifies and removes dynamic distractors.
• A new benchmark dataset of drone-captured scenes (dynamic and static) was created for evaluation.

Main Results:

• DroneSplat+ achieves accurate geometry recovery even with limited drone viewpoints.
• The method successfully suppresses dynamic elements, preserving static scene integrity.
• Experimental results demonstrate superior performance compared to 3D Gaussian Splatting (3DGS) and Neural Radiance Fields (NeRF) baselines.

Conclusions:

• DroneSplat+ provides a robust solution for 3D reconstruction from challenging in-the-wild drone imagery.
• The semantic-driven approach effectively resolves the interplay between limited views and dynamic scenes.
• The proposed framework and dataset advance the field of drone-based 3D scene understanding.