GeoSemField

While 3D semantic Gaussian splatting pipelines establish scene-level semantic associations, they inherently fail to translate this continuous relevance into precise target-extent prediction. Existing language-conditioned pipelines typically render a semantic feature map and derive predictions through post-hoc similarity comparisons with a text query. While these methods effectively highlight semantically relevant neighborhoods, they fail to definitively assign rendered pixels or Gaussians to a queried target when objects are adjacent, nested, or frequently co-occurring. We investigate this critical gap between semantic relevance and true target membership, proposing \textbf{GeoSemField}, a geometry-aware and language-conditioned semantic field tailored for explicit target-extent prediction in 3DGS. Our framework extracts viewpoint-aligned semantic carriers from a shared 3DGS representation and optimizes a query-conditioned target field over this rendered evidence. By incorporating feature-space Eikonal regularization alongside geometry-alignment constraints, our method strictly regulates response transitions. This formulation ensures that continuous semantic relevance is reliably converted into stable, discrete target boundaries. Extensive experiments on open-vocabulary 3D segmentation and localization benchmarks, supported by comprehensive ablations and leakage diagnostics, demonstrate that explicit target-membership estimation significantly enhances object isolation and mitigates semantic leakage in cluttered environments compared to existing methods.

Dataset

Data preparation

All datasets can be obtained by following the official instructions or download links provided by the corresponding baseline projects:

• LERF-OVS: please follow the dataset preparation instructions in LangSplat
• 3D-OVS: please follow the official release in 3D-OVS
• MipNeRF360v2: please download the data from the official Mip-NeRF 360 page

Train

Directory Layout

standard_code/
├── configs/fusion_mainline.yaml
├── fusion/
│   ├── run_fusion_training.py
└── geosemfield/
    ├── arguments/
    ├── gaussian_renderer/
    ├── scene/
    ├── utils/
    ├── geosemfield/
    ├── tools/
    ├── ext/
    └── submodules/

Environment

Create the environment from environment.yml, then build the local CUDA extensions:

conda env create -f environment.yml
conda activate goi
cd geosemfield/submodules/diff-gaussian-rasterization
pip install -e .
cd ../simple-knn
pip install -e .
cd ../../..