---
license: other
tags:
  - 3d-gaussian-splatting
  - open-vocabulary-segmentation
  - clip
  - sam
---

# THGS — LERF-OVS language_features (precomputed)

Precomputed per-view language features for the 4 LERF-OVS scenes
(`figurines`, `ramen`, `teatime`, `waldo_kitchen`), used as input to the
[THGS](https://github.com/BAEJUNHAK/THGS) pipeline (`merge_proj.py`, Stage 3 replay).

For each training image there are two files:

| file | content |
|---|---|
| `frame_XXXXX_s.npy` | per-view SAM segmentation maps (4-level, LangSplat-variant SAM) |
| `frame_XXXXX_f.npy` | per-mask CLIP features |

Generated with `scripts/image_encoding.py` using the
[segment-anything-langsplat](https://github.com/minghanqin/segment-anything-langsplat)
SAM variant (the standard `segment-anything` does **not** work — `generate()`
must return a 4-level tuple) and `sam_vit_h_4b8939.pth`.
Regenerating from scratch takes ~4 GPU-hours; this dataset lets you skip that.

## Usage

Download into your THGS repo so the layout becomes
`data/lerf_ovs/<scene>/language_features/*.npy`:

```bash
hf download JUNHAKBAE/THGS-lerf-ovs-language-features \
    --repo-type dataset --local-dir data/lerf_ovs
```

Total size ≈ 9 GB (figurines 3.4G, ramen 1.5G, teatime 2.0G, waldo_kitchen 2.1G).