Tac-DINO / Scan3dTac

Upload complete. All archives have passed remote SHA-256 verification. Final indexes and exact deduplication statistics are available in release_manifest.json and checksums.json.

The author's Scan3dTac data contains 96 categories and 494 organized objects: 347 train and 147 test, with no object overlap. There are 19,619 tactile videos and 9,386 processed RGB views. The material map contains 15 labels.

Release structure

• raw/assets/: Scan3dTac's original 3D_Scanning/ scanner assets, direct Tactile/ RGB/tactile videos and depth HDF5 files, plus Annotation/.
• Scan3dTac/core/: annotations, tactile MP4 clips and Preprocessed0111/.
• Scan3dTac/geometry/: tactile point clouds and contact geometry.
• Each group's metadata/files.jsonl.gz maps original relative paths to unique SHA-256 payloads and archive names. core.files.jsonl.gz is the core-only index.

Content is deduplicated separately within raw and Scan3dTac: a given SHA-256 payload is archived exactly once per group. Repeated logical paths refer to the same stored payload. This preserves original frame sequences, filenames and calibration files without storing duplicate bytes. Corresponding content may appear once in each of the two groups.

Only the 494-object Scan3dTac collection is included. ObjectFolderReal, transfer archive copies and unrelated datasets are excluded. Source files have not been modified. The processed release uses the current Preprocessed0111/ version and derived geometry. The older Preprocessed/ directory is excluded from this release.

Both raw and processed shards use TAR + Zstandard level 3 with frame checksums. The choice was benchmarked on Scan3dTac depth/geometry data, where Zstandard compressed better and faster than gzip. Shards target 4 GiB of unique uncompressed payloads, with a smaller first shard for transfer verification. The final source counts, deduplication savings and planned archives are in release_manifest.json.

Download and restore

Use the standalone downloader to fetch only the required archives and restore ordinary files under their original names. For the complete training core:

pip install "huggingface_hub>=0.32,<3" "zstandard>=0.23"
curl -L https://huggingface.co/datasets/lihong-cs/Tac-DINO/resolve/main/download_dataset.py -o download_dataset.py
python download_dataset.py --group Scan3dTac --profile core --output ./data/scan3dtac

The core is already available: 13 archives / 16.25 GB compressed, restoring the 48.05 GB training/reprocessing subset. Allow approximately 55 GB of free space for this selection. Configurations requiring per-contact 3D inputs also need the geometry group; the no-3D configurations can use the core alone.

Other selections, once their required archives are published:

# Complete processed collection (core plus geometry).
python download_dataset.py --group Scan3dTac --output ./data/scan3dtac
# Add geometry to an existing core download.
python download_dataset.py --group Scan3dTac --profile geometry --output ./data/scan3dtac
# Original Scan3dTac assets.
python download_dataset.py --group raw --output ./data/raw
# One object for inspection/postprocessing.
python download_dataset.py --group Scan3dTac --profile core --object wipes/obj_0000 --output ./data/example
# Restrict to an official split or category; --object and --category are repeatable.
python download_dataset.py --group Scan3dTac --profile core --split train --category bowl --output ./data/bowls
# Inspect availability and sizes without downloading payload archives.
python download_dataset.py --group raw --output ./data/raw --dry-run

The downloader pins a repository commit, checks archive and payload SHA-256, restores one archive at a time, then deletes its compressed download cache. Rerun the same command to resume. Completed files are tracked outside the dataset directory; --verify-existing additionally rehashes them. Only the final restored files plus one compressed archive are needed, without an extra uncompressed hash/blob tree. Disk space is checked before each archive. On unstable connections, --http-chunk-mb 1 uses smaller resumable HTTP requests.

Original MP4, HDF5, NPY/NPZ, PNG/EXR, GLB/PLY/OBJ and JSON bytes are preserved. The default creates independent ordinary files, so editing an annotation does not change another object with identical content. --hardlinks is optional for immutable datasets; linked aliases share subsequent edits. Label maps, original train/test lists and selection records are saved beside the output, for example ./data/scan3dtac_metadata/. Tac-DINO's extra can point to the restored dataset and its metadata root to that companion directory.

Archive-level selection may download extra bytes for other objects in the same archive, but only selected original files are retained. While publication is in progress, the tool reports missing required archives before downloading payloads. An optional download mirror can be used with --endpoint https://hf-mirror.com; archive and payload checksums are still checked. The public download requires no Hugging Face token. An existing official-Hub login is used when available; --anonymous disables it, and mirrors use anonymous requests for this public dataset. Resumable HTTP is the default transport for proxy/mirror compatibility; --use-xet enables the native Xet downloader when preferred. Already-downloaded repositories can be read with --repo-dir.

The download/restore workflow was checked against the original collection using the 237-file wipes/obj_0000 core sample and one raw depth HDF5 file. Every restored file matched its original SHA-256, and a repeat run reused the completed restoration without downloading the archive again.

The older restore_dataset.py remains available for offline restoration.

This release supports local training and postprocessing after restoration. The Hub viewer shows an object index; the archived payloads are not a native load_dataset() streaming configuration. Use the supplied object split and matching Tac-DINO loader; do not randomly split views of one object across train and test.

Provenance and documentation

Raw and processed indexes preserve the same <category>/<object>/... IDs. Before content deduplication or compression, raw assets contain 81,534 files / 1,146,904,479,398 bytes; processed data contains 191,913 files / 207,660,168,056 bytes. Annotations and training tactile clips are retained once in each group so either group can be restored independently. Coordinate transforms, physical units, depth conventions, calibration and capture hardware require additional documentation from the collector.

License and citation

The data license and formal citation are pending the collector's selection. Model/code licenses do not automatically apply to the collected dataset. ObjectFolderReal, checkpoints, logs and third-party code are outside this release.