Skip to content

DATA.md

Latest commit

 

History

History
50 lines (36 loc) · 1.58 KB

DATA.md

File metadata and controls

50 lines (36 loc) · 1.58 KB
Raw

Data Preparation

Defaults in tokengs/data/registry.py resolve dataset roots under the repository root:

Path Role
data/dl3dv Training zips for DL3DV10K (e.g. DL3DV-ALL 960p undistorted).
data/dl3dv_eval Eval set for DL3DVEval (e.g. DL3DV-10K-Benchmark).
data/kubric Kubric multi-view 4D tar dump used by finetune_dl3dv_kubric_* presets.

Implementation details for readers and transforms live in tokengs/data/static/dl3dv.py.

Symlinks (recommended)

From the repository root:

mkdir -p data
ln -snf /absolute/path/to/DL3DV-ALL-960P-undistorted data/dl3dv
ln -snf /absolute/path/to/DL3DV-10K-Benchmark data/dl3dv_eval
ln -snf /absolute/path/to/objaverse_4d/kubric_mv data/kubric

-snf creates or replaces a symlink. Relative targets (e.g. ../datasets/dl3dv) are fine if paths stay stable.

Kubric Layout

The Kubric reader expects split directories and per-camera tar files:

data/kubric/
  v0/
    <scene>/
      output_000.tar
      output_001.tar
      ...
  v1/
  v2/

Each output_{view:03d}.tar should contain metadata.json, rgba_{frame:05d}.png, and depth_{frame:05d}.tiff. The dynamic presets enable pointmap camera scaling, so the depth TIFFs are loaded for the input frames.

Overrides without symlinks

Pass kwargs the dataset constructor accepts (for example root_path) via Tyro. See dataset_kwargs on Options and run:

python -m tokengs.train --help
python -m tokengs.evaluate --help