Najongs/dinobotpose-datasets

DINObotPose — own-capture datasets

Multi-view captures recorded for DINObotPose, packed as zstd tarballs. Start with the tarballs. The loose ICRA_multiview/ tree is the remnant of an interrupted earlier upload — see Repository layout below.

Reproducing the paper does not require this repo: scripts/prepare_dream.py in the release repo builds what it needs from the public DREAM dataset. This repo holds the parts DREAM does not contain.

Tarballs

File	Size	Contents
franka_research3.tar.zst	9.9 GiB	Franka Research 3, multi-view capture
misc_datasets.tar.zst	9.7 GiB	meca500_synth · meca_insertion · meca500 · Fr5_label_output · captures · FR5_robot · ZED
Fr5.tar.zst	3.2 GiB	FR5, multi-view capture
Meca_insertion.tar.zst	1.2 GiB	Meca500 needle-insertion capture
Meca500.tar.zst	200 MiB	Meca500 capture
Converted_own.tar.zst	17 MiB	DREAM-format conversions of the above, plus the train/val split symlinks

Tarballs total ≈ 24.2 GiB, packed from 48 GB of loose files (the JSON labels compress well).

hf download Najongs/dinobotpose-datasets --repo-type dataset --include "*.tar.zst" --local-dir .
for f in *.tar.zst; do tar --zstd -xf "$f"; done

What is deliberately not here

DREAM_syn (Panda synthetic) and the pristine synthetic/ splits are not redistributed — they are NVIDIA's DREAM data, re-obtainable via DREAM/data/DOWNLOAD.sh. Measured on the source tree, DREAM originals and their conversions were 88% of the file count; only the 6% that is own-capture is kept here.

The conversion and preprocessing scripts live in the code repo, not here.

Repository layout — why there are loose files too

ICRA_multiview/ is left over from a first upload attempt that hit the Hugging Face free-tier limit (1,000 API requests per 5 minutes) partway through 708,342 files; 23,574 of them landed. Packing into 6 tarballs is what made the upload feasible. The remnant is kept rather than deleted because deleting does not reclaim storage on the Hub, and part of it is not duplicated by any tarball:

Path	Also in a tarball?
ICRA_multiview/Fr5/	yes — Fr5.tar.zst
ICRA_multiview/DREAM_real/	no — DREAM data, re-obtainable
ICRA_multiview/*.py (6 files)	no — preprocessing/calibration scripts, now tracked in the code repo

Original DREAM data is by its own authors under its own terms; this repo redistributes derived splits for reproducibility only.

---

Jongyeol Na — Researcher, KIRO · M.S. DGIST, IROM Lab GitHub · Knowledge graph · W&B