Feature fields
Distilled feature field: a 3D function trained so its rendering from each camera matches a 2D teacher's feature map.
| Work | Key fact | Versus this project |
|---|---|---|
| Distilled Feature Fields, LERF | CLIP, DINO or LSeg features distilled into a NeRF; text, patch or click queries | Implicit, 3D, per scene |
| Feature 3DGS | Arbitrary-dimension features on 3D Gaussians (SAM, LSeg) | 3D, per scene, optimised features |
| LangSplat | CLIP features via a scene-wise autoencoder; 199x faster than LERF at 1440x1080 | Compresses features; yours keeps 384-d |
| Upsample Anything | One Gaussian kernel per coarse pixel; 0.419 s per 224 px image | Filter on the grid; features not on Gaussians |
| This project | 2D Gaussians own features; closed-form solve; amortised placement | Single image, any count |
Defence answer. “Feature 3DGS puts optimised features on 3D Gaussians per scene. I put closed-form features on 2D Gaussians placed by a trained network for one image, at any count. Upsamplers produce a dense grid; mine is a compact set read at any location. A matched-protocol comparison with upsamplers is future work.”
Caveats: semantics are inherited from the teacher, which is still needed at inference; no fair upsampler comparison yet.