starting_kit/README.md

# TwinWorld 2026 Starting Kit

This page covers two things: the submission format, and a set of reference
implementations you can use as a starting point.

## Submission Format

A single zip file, containing the dataset folders at the top level:

```
submission.zip
├── tum/<scene_id>/
│   ├── rgb/<frame_id>.png
│   └── 3D_point_cloud/point_cloud.ply         # x, y, z
└── gold_coast/<scene_id>/
    ├── rgb/<frame_id>.png
    └── 3D_point_cloud/point_cloud.ply         # x, y, z, classification
```

- Scene IDs and frame IDs (file name stems) must exactly match the reference frames provided for each phase.
- rgb/ images must be PNG at exactly the reference resolution.
- Point clouds must be valid PLY with finite x, y, z coordinates. Binary little-endian PLY with float32 coordinates is recommended to keep uploads small; store only the properties listed above (no colors/normals — they are ignored by the scorer and inflate the upload).
- Gold Coast point clouds additionally require an integer classification property with values in {0, 1, 2, 3, 4, 255}.

Note: see the CodaBench challenge page for the complete and authoritative submission rules.

## Reference Implementations

These are starting points, not requirements. Any method that can be trained
on this challenge's COLMAP-format data and produce the outputs above is a
valid submission.

### Basic Gaussian reconstruction

- **Original 3D Gaussian Splatting**
  Code: https://github.com/graphdeco-inria/gaussian-splatting

- **gsplat**
  Code: https://github.com/nerfstudio-project/gsplat
  This is the Gaussian Splatting rasterization backend maintained by the
  Nerfstudio team, and it is what Nerfstudio's own `splatfacto` method is
  built on. If you are already comfortable with the Nerfstudio ecosystem
  (data parsers, config system, viewer), this lets you stay in that workflow
  instead of adapting the original repo above.

### Geometry-oriented Gaussian reconstruction

For the TUM geometry track, these methods target more accurate surfaces
than vanilla 3DGS, which should help the F-score and Chamfer distance
metrics.

- **2D Gaussian Splatting**
  Code: https://github.com/hbb1/2d-gaussian-splatting

- **PGSR**
  Code: https://github.com/zju3dv/PGSR

Note: these methods represent the scene with different primitives than
vanilla 3DGS (oriented disks for 2DGS, planar Gaussians for PGSR). Most of
the semantic methods below are built as extensions of the original 3DGS
codebase and assume that representation, so they cannot simply be bolted
on top of a 2DGS or PGSR scene without adaptation.

### Semantic Gaussian representation

For the Gold Coast semantic track:

- **Gaussian Grouping**
  Code: https://github.com/lkeab/gaussian-grouping

- **Gaga**
  Code: https://github.com/weijielyu/Gaga

- **OpenGaussian**
  Code: https://github.com/yanmin-wu/OpenGaussian

- **LangSplat**
  Code: https://github.com/minghanqin/LangSplat

Note: none of these methods directly output our required `classification`
values, and the extra step needed differs by method.

- Gaussian Grouping and Gaga group Gaussians using 2D masks (for example from
  SAM) into class-agnostic instances: the result is a consistent object ID
  per Gaussian, not a semantic class name, so you still need to map each
  group to one of our 5 classes yourself.
- LangSplat and OpenGaussian instead train a further per-Gaussian feature on
  top of an already-reconstructed scene and let you query open-vocabulary
  text prompts: you still need an extra step to turn that similarity score
  into one discrete class ID per point, using our 5 class names as the
  prompts.

In both cases, turning the method's native output into our required PLY
`classification` property is left to you.