Skip to content

Latest commit

 

History

History
34 lines (24 loc) · 1.71 KB

README.md

File metadata and controls

34 lines (24 loc) · 1.71 KB
Raw

[CVPR2024 Highlight] LangSplat: 3D Language Gaussian Splatting

  • release more preprocessed dataset and the pretrained model (coming soon)
  • release the code of the eval

This project is still under development. Please feel free to raise issues or submit pull requests to contribute to our codebase.

branch explanation:

  • prev_generate_maps: generate 3d query sim maps and render them

    The main changes happened in the eval folder and gaussian_renderer folder. And the render.py file

  • eval_only: only eval the 3d query's output, not 2d

  • langplat_vq: generate vq codebook and indices

image image

Attention This repo has frequent hard codes and duplicate codes. For high-level learning, understanding, and semi-test only. DON'T TRY TO RUN IT!!!!

Usage:(Try not to run the code.) without vq:

  • prev_generate_maps:

    1. use _mytest.py to generate language_feats_dim3_tensor_1.pt, and copy them to the eval/ folder.

    2. go to eval folder, sh eval.sh to generate 0_valid_map_level_1_neg_0.pt, etc. This is the sim map, show be 1 pos and 4 neg

    3. render:

      3.1 go back to base folder, edit the render.py file, hard code exist here. Modify the corresponding parts at meta_data.info. Then sh process.sh

      3.2 hardcode exist here - LangSplat/gaussian_renderer/init.py. modify it

    4. output should be LangSplat/dataset/lerf_ovs/figurines/output/figurines_1/train.

  • eval_only

    1. only the eval folder changed. Hardcode here.