Fashionpedia is a COCO-style fashion understanding dataset that combines instance annotations (object/box/segmentation) with attributes and an explicit category–part–attribute ontology. Its key contribution is enabling part-level attribute localization (e.g., “the sleeve is striped”) within a unified benchmark for detection, segmentation, and attribute prediction.
Earlier fashion datasets typically emphasized either categories or attributes, but did not explicitly model the relationship between garment parts and attributes. Fashionpedia addresses this by providing a structured ontology and detailed part-aware annotations.
| Component | Description |
|---|---|
| Categories | 27 high-level clothing categories |
| Parts | 19 clothing part labels (e.g., collar, sleeve) |
| Attributes | 46 attribute tags (e.g., solid, floral, knit) |
| Ontology | Hierarchical links between category–part–attribute |
| Masks | Pixel-level segmentation masks for garments and parts |
| Attributes (part-level) | Attributes annotated at the part level |
The dataset is distributed as COCO-style JSON annotations and supports multi-task learning and ontology-aware evaluation.
fashionpedia_root/
├── instances_attributes_train2020.json (core)
├── instances_attributes_val2020.json (core)
├── attributes_train2020.json (optional)
├── attributes_val2020.json (optional)
├── info_test2020.json (optional)
└── images/ (optional)
├── train/
├── val/
└── test/The FEL extractor normalizes Fashionpedia into ontology (terms), observation (instances/labels), and audit (manifest/QC) layers.
fp_terms_categories.csv (category + part terms; includes term_role)fp_terms_attributes.csv (attribute terms)fp_images_index.csv.gz (ImageItem hub keyed by image_uid)fp_instances.csv.gz (InstanceItem core table)fp_attr_sparse.csv.gz (image-level attribute sparse; separated to avoid confusion)fp_image_categories.csv.gz and fp_image_categories_agg.csv.gz (derived category log + aggregation)fp_geometry.csv.gz (derived geometry summary per image)items.csv.gz (typed registry for emitted entities)manifest.csv / manifest.jsonlqc_summary.jsonreport.mdFashionpedia strengthens FEL’s InstanceItem layer: meaning is expressed at the instance level (garment/part) with category + attributes + spatial evidence.
image_uid from fp_images_index.csv.gzinstance_uid from fp_instances.csv.gzfp_cat_id from fp_terms_categories.csvfp_attr_id from fp_terms_attributes.csvimage_uid from fp_geometry.csv.gz
ImageItem ─contains────────▶ InstanceItem
InstanceItem ─has_category──▶ CategoryTerm
InstanceItem ─has_attribute─▶ AttributeTerm
ImageItem ─has_geometry────▶ GeometrySummary
(Separate) ImageItem ─has_attribute──▶ AttributeTerm (image-level sparse; fp_attr_sparse.csv.gz)
Fashionpedia extraction is driven by the COCO-style JSONs; there is no benchmark folder split like DF1. Instead, the normalized schema supports multiple tasks (detection/segmentation/attributes) from one consistent graph layer.
COCO ann.id is not globally unique across JSONs (e.g., train vs val), so it is not safe as a primary key.
FEL v1.3 solution (deterministic, globally unique):
instance_uid = FP:inst/<split>/<src_file_rel>/<row_index>This ensures 100% uniqueness via split + dataset-relative provenance + JSON row position.
The Fashionpedia normalized graph is image-hub with strong instance-level semantics.
fp_images_index.csv.gz is the hub keyed by image_uid.fp_instances.csv.gz): instance_uid + bbox + seg summary + instance-level attributesfp_attr_sparse.csv.gz): image-level sparse attributes (kept separate)fp_terms_categories.csv) lookup by fp_cat_idfp_terms_attributes.csv) lookup by fp_attr_idfp_geometry.csv.gz): derived geometry summary per image_uiditems.csv.gz) is a lightweight typed registry abstraction.Based on the finalized v1.3 extractor policy, the following checks define “artifact correctness” for FEL ingestion:
items.csv.gz must be a subset of emitted fp_instances.instance_uid.fp_instances.src_file is POSIX, dataset-relative, and is embedded into instance_uid.instance_uid duplicates are hard-fail; instance_id duplicates are allowed but counted/reported.Expected QC outcomes under this policy:
instance_uid duplicate = 0 (hard-fail if > 0)hard_fail = False when required entities exist and PK/integrity conditions holdinstance_uid.Under the stated FEL v1.3 QC policy, a “submittable” run satisfies:
instance_uid duplicate = 0categories > 0, attributes > 0, images > 0, instances_emitted > 0hard_fail = False➡️ This indicates PK stability, referential integrity, ontology separation, and auditability are all satisfied for FEL ingestion.
Click below to open the interactive graph in a new window:
Open Fashionpedia Graph Interactive Editorinstance_uid eliminates COCO ID collisions across splits and files.➡️ Core FEL input for ontology-aware segmentation, part-level attribute localization, and instance-level reasoning across fashion datasets.