OutLangSplat: 3D Language Gaussian Splatting for UAV Outdoor Scenes

OutLangSplat is a new 3D language Gaussian splatting approach aimed at UAV outdoor scenes, where open-vocabulary scene understanding has been much less reliable than in indoor or small-scale settings. The core problem is familiar to anyone working with sparse viewpoints: distant objects, heavy occlusion, and noisy views can trigger the wrong semantics or leave targets partially unresolved.

To address that, the system uses a 2D-3D dual-branch representation with region-based alignment and fusion. In practical terms, that is meant to keep spatial semantics more consistent across views, reducing both incomplete target responses and background activations that can poison downstream segmentation or localization.

The second piece is a training-free Gaussian feature aggregation strategy that weighs pixel contribution reliability and cross-view semantic consistency. That matters because it avoids adding another learned module just to clean up noisy viewpoints, which can be attractive for teams trying to keep pipelines simpler and easier to reproduce.

The project also includes a manually annotated dataset built from four real-world public UAV outdoor scene datasets, described as the first accessible open-vocabulary 3D scene understanding dataset for this setting. Quantitative results and ablations reportedly show gains over current SOTA on open-vocabulary semantic segmentation and instance localization, and the code and data are planned for release. For graphics and tools teams, the bigger takeaway is that language-grounded 3D representations are starting to look viable outside...