Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

MolmoMotion-1M — dataset construction

The MolmoMotion-1M training corpus — including every per-dataset reconstruction recipe — lives on HuggingFace:

allenai/molmo-motion-1m

For each of the seven sub-datasets (egodex, ytvis, hdepic, xperience, stereo4d, droid, molmospaces) the dataset repo ships:

  • README.mdauthoritative schema + reconstruction procedure for that dataset
  • annotations/ — clips / split / index JSONs (the source of truth)
  • tracks/ — the 3D (and, for most datasets, 2D) point tracks we produced (track_index/ for stereo4d, which reconstructs its tracks locally)
  • camera/ — per-frame pose + intrinsics (absent for xperience, which reconstructs it)
  • videos/ — shipped only for molmospaces; every other dataset reconstructs videos locally from the original upstream source via the bundled reconstruct_* script, due to license restrictions

Download:

huggingface-cli download allenai/molmo-motion-1m \
    --repo-type dataset --local-dir $MOLMO_MOTION_1M_ROOT

Then follow each <dataset>/README.md inside the downloaded root to reconstruct the videos (and the per-dataset remaining signals). Nothing in this directory is needed for training — point MOLMO_MOTION_1M_ROOT at the downloaded corpus and go.