chibifire/zenodo-second-hand-fashion-v3
Second-Hand Fashion Dataset — wide (one row per garment) Repack of Zenodo record 10.5281/zenodo.13788681 (Nauman et al., RISE + Wargön Innovation + Myrorna, CC-BY-4.0) into a one-row-per-garment wide layout so the HF dataset viewer shows every attribute — three images plus 25 metadata columns — on a single row. Previous v3 releases stored one row per (garment, view) with satellite tables that had to be joined manually. That layout is preserved in git history if you need it; the… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/zenodo-second-hand-fashion-v3.
Second-Hand Fashion Dataset — wide (one row per garment)
Repack of Zenodo record 10.5281/zenodo.13788681 (Nauman et al., RISE + Wargön Innovation + Myrorna, CC-BY-4.0) into a one-row-per-garment wide layout so the HF dataset viewer shows every attribute — three images plus 25 metadata columns — on a single row.
Previous v3 releases stored one row per (garment, view) with satellite tables that had to be joined manually. That layout is preserved in git history if you need it; the current release inlines everything.
Splits
Split boundary is the garment: all views for one garment_id belong to the same split.
- eval is the upstream gold set (
is_gold == 1) — small, hand-curated, appropriate for development inspection. - test is a blinded 10% holdout carved from non-gold garments with a fixed seed. Per this workspace's holdout discipline (patterned on COCO
val2017), it is not inspected during dev, not used to pick a checkpoint or hyperparameter or threshold, and not used as a generator source. - train is everything else.
Row schema
Each row is one garment. 28 columns; up to three JPEG image columns; all categorical *_id fields resolved to human-readable strings; multi-valued attributes (colors, damage) stored as list columns.
Layout
data/{train,test,eval}-*.parquet — one row per garment, ZSTD-compressed.
Named-and-counted skips (rule 3)
From build_manifest.json and resplit_manifest.json:
- Garments scanned: 32,036 (31,638 upstream base set + 398 gold
test100label files). - Distinct gold garment_ids that yielded any image row: 298 (100 gold label files pointed at shared images already covered by other gold labels).
- Malformed JSON: 0.
- Missing brand images: 5,290 total — matches the upstream README's ~4,000-5,000 removed for privacy or absent because the garment carried no brand label. In the wide layout these appear as
image_brand = null.
Legacy normalized artifacts
The original ETNF satellite/vocab tables (satellite_*.parquet, vocab_*.parquet, garment_lookup.parquet) are retained at the repo root for anyone who needs the normalized keying — the wide data/ shards are a full denormalized snapshot of them joined on garment_id.
No derived columns
condition, pilling, etc. are stored as-is; nothing implies is_recommended or similar. Free-text damage was not backfilled from the damage-slot enums.
Source DOI: 10.5281/zenodo.13788681. See CITATION.cff for full attribution.