Ryan A. / Computer vision research

vgrep3d

Search a Gaussian splat scene with natural language.

A research pipeline for text-guided 3D scene understanding using SigLIP 2 semantic features, compressed per scene and distilled into a Gaussian feature field.

Research status: partially working. Semantic heatmaps and broad localization are demonstrated. Tight object boxes and instance separation remain unresolved. This Space presents the project; it does not run GPU inference.

From images to searchable 3D

Posed images → Gaussian scene → semantic feature extraction → compressed feature field → text query.

Explore the experiments

The repository includes visual results, setup instructions, evaluation guidance, and documented failure cases.