Research map / 3D vision / Gaussian splatting
DroneSplat: 3D Gaussian Splatting for Robust 3D Reconstruction from In-the-Wild Drone Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DroneSplat, a 3DGS framework for in-the-wild drone imagery that adaptively masks dynamic distractors via local-global SAM v2 segmentation with residual statistics, and improves limited-view geometry with DUSt3R-based point sampling and voxel-guided optimization.
Problem
3D reconstruction from in-the-wild drone imagery is hurt by dynamic distractors that violate static-scene assumptions and by sparse viewpoints that cause unconstrained 3DGS optimization and poor novel-view quality.
Approach
Adaptive Local-Global Masking combines per-object residual averages from SAM v2 segments, a statistically set adaptive threshold (expectation plus variance), and SAM v2 video tracking of high-residual candidates; for sparse views, DUSt3R dense points are confidence/FPFH-sampled per voxel and 3DGS is optimized with voxel-bounded Gaussians.
Key results
- On the DroneSplat dynamic scenes, Ours reaches 20.86 PSNR / 0.618 SSIM / 0.217 LPIPS on average-level ablation versus 19.02 for the baseline 3DGS-style ablation (a), and outperforms all baselines per scene.
- On the DroneSplat static limited-view set, Ours achieves 18.09 PSNR (Exhibition Hall) and 18.17 PSNR (Plaza), above InstantSplat (17.37, 16.61) and Scaffold-GS (16.55, 17.01).
- On the On-the-go dataset ablation, the full model reaches 22.77 PSNR versus 18.46 for the baseline without masking or voxel guidance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DroneSplat rather than the strongest baseline for distractor removal | DroneSplat · DroneSplat dynamic scenes | PSNR | Mixed |
| Use DroneSplat rather than the strongest baseline for distractor removal | DroneSplat · NeRF On-the-go | PSNR | Mixed |
| Use DroneSplat rather than the strongest baseline for limited-view reconstruction | DroneSplat · DroneSplat static and UrbanScene3D | PSNR | Helped |
| Add adaptive local masking (ablation) | DroneSplat with vanilla 3DGS · DroneSplat(dynamic) | PSNR: 19.02 -> 20.71 | Helped |
| Add adaptive local masking (ablation) | DroneSplat with vanilla 3DGS · NeRF On-the-go | PSNR: 18.46 -> 22.64 | Helped |
| Add complementary global masking after local masking (ablation) | DroneSplat with adaptive local masking · DroneSplat(dynamic) | PSNR: 20.71 -> 20.74 | Helped |
| Add complementary global masking after local masking (ablation) | DroneSplat with adaptive local masking · NeRF On-the-go | PSNR: 22.64 -> 22.71 | Helped |
| Add voxel-guided 3DGS after local and global masking (ablation) | DroneSplat · DroneSplat(dynamic) | PSNR: 20.74 -> 20.86 | Helped |
| Add voxel-guided 3DGS after local and global masking (ablation) | DroneSplat · NeRF On-the-go | PSNR: 22.71 -> 22.77 | Helped |
| Add geometric priors without point sampling or Gaussian optimization (ablation) | 3DGS · DroneSplat(static) | PSNR: 16.01 -> 16.99 | Helped |
| Add geometric priors without point sampling or Gaussian optimization (ablation) | 3DGS · UrbanScene3D | PSNR: 16.53 -> 15.33 | Hurt |
| Add geometric-aware point sampling to geometric priors (ablation) | 3DGS · DroneSplat(static) | PSNR: 16.99 -> 17.46 | Helped |
Surprising findings
- Vanilla 3DGS initialized with multi-view stereo points does not improve reconstruction quality, indicating unconstrained optimization undermines the geometric priors.
- Hard-coded residual thresholds are reported to be tedious to tune and fixed during training, which the adaptive threshold is meant to address.
Limitations
- Evaluation on limited-view scenes uses only 6 input views and a small number of scenes (2 static, 2 UrbanScene3D).
- Some metrics are not best across all scenes; for example, NeRF-HuGS and GS-W beat the method on some individual sequences.
Experimental setup
Datasets: DroneSplat dataset (24 drone sequences) On-the-go dataset UrbanScene3D
Benchmarks: DroneSplat dataset On-the-go dataset UrbanScene3D
Metrics: PSNR SSIM LPIPS
Baselines: RobustNeRF NeRF-HuGS NeRF On-the-go GS-W WildGaussians Mip-Splatting FreeNeRF DiffusioNeRF FSGS DNGaussian InstantSplat Scaffold-GS
Builds on: 3D Gaussian Splatting RobustNeRF DUSt3R InstantSplat Segment Anything Model v2
Closest papers by meaning
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.87)
- Wild-GS: Real-Time Novel View Synthesis from Unconstrained Photo Collections (NeurIPS 2024 · similarity 0.87)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.86)
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.86)
- WildGaussians: 3D Gaussian Splatting In the Wild (NeurIPS 2024 · similarity 0.86)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.