DeepFashion (DF1) — Normalized Dataset for FEL

1. Dataset Overview

DeepFashion (DF1) is a large-scale fashion image dataset designed for clothing recognition and retrieval research.
It reflects real-world conditions such as pose variation, deformation, occlusion, and lighting.

In this pipeline, the original DF1 dataset has been restructured into a normalized schema for use in the
FEL (Feature Evidence Layer or Knowledge-Graph-based Learning Layer).

Key Characteristics

• Over 800,000 images with rich annotations (categories, attributes, landmarks, etc.)
• Composed of five benchmarks:
  • In-shop Retrieval
  • Consumer-to-shop Retrieval
  • Category & Attribute Prediction (CAPB)
  • Landmark Detection
• All data unified using a single join key: image_uid
• Multimodal information stored in separate tables:
  • Text
  • Spatial data (bounding boxes)
  • Structural data (landmarks)
  • Labels (category/attribute)
  • Evaluation pairs
• Provenance preserved (src_file + row_index) for reproducibility

➡️ This converts DF1’s heterogeneous formats into a standardized schema optimized for multimodal learning and graph-based analysis.

2. Folder and File Structure

(1) Original DF1 Structure (Input)

Evaluation Splits

• list_eval_partition.txt → Train / validation / test or query / gallery partition

Bounding Box Files

• list_bbox_inshop.txt
• list_bbox_consumer2shop.txt
• list_bbox.txt → Clothing object location

Landmark Files

• list_landmarks.txt → Keypoint coordinates

Category / Attribute Files

• list_category_*
• list_attr_* → Taxonomy definitions and labels

Text Descriptions

• list_description_inshop.json → Natural-language descriptions

(2) Normalized Output Structure

Each benchmark is converted into:

benchmarks/ df1inshop/normalized/ df1consumer2shop/normalized/ df1capbcoarse/normalized/ df1capbfine/normalized/ df1_landmark/normalized/

Core Data Tables

• eval_partition.csv.gz
• bbox.csv.gz
• landmarks_raw.csv.gz
• category_labels.csv.gz
• attr_sparse.csv.gz
• retrieval_pairs.csv.gz
• descriptions.csv.gz
• image_sources.csv.gz

Ontology / Term Tables

• category_terms.csv.gz
• attr_terms.csv.gz

Management & Validation

• manifest.csv / manifest.jsonl
• qc_summary.json

3. Role in FEL

Node Construction

• Image → image_uid
• Category → category_terms
• Attribute → attr_terms
• Text → descriptions
• Optional Item → bbox/landmark derived

Relationship Construction

Image ─hasbbox──────────▶ BBoxEvidence
Image ─haslandmark──────▶ LandmarkEvidence
Image ─hascategory──────▶ CategoryTerm
Image ─hasattribute────▶ AttributeTerm
Image ─has_description──▶ Text
Image(query) ─matches───▶ Image(gallery)

Core Design Principles

Single Join Key
All data linked via image_uid

Evidence Separation
Ontology / Labels / Evidence split

Provenance Preservation
src_file + row_index stored

4. Extracted Benchmarks

In-shop Retrieval

• Images: 52,712
• Text descriptions: 8,081
• No bbox/landmark errors

Consumer-to-Shop Retrieval

• Images: 239,557
• Retrieval pairs: 195,540
• Real shopping scenario matching

CAPB

Coarse: - Images: 289,222 - Categories: 50 - Attributes: 1,000

Fine: - Images: 20,000 - Attributes: 26

Landmark Benchmark

• Images: 123,016
• No landmark errors

Excluded Data

Original image files not included (metadata‑focused FEL design).

5. Graph Structure Description

Central Node

Images — anchor for all connections

Data Ingestion

ImageSources → Images (canonicalization)

Labels & Attributes

CategoryTerms → CategoryLabels → Images
AttrTerms → AttrSparse → Images
ColorTerms → ColorLabels → Images

Evaluation Structure

• EvalPartition → dataset splits
• RetrievalPairs → retrieval evaluation
• Descriptions → text data

Reproducibility

Manifest → artifact audit
QCSummary → pipeline validation

DF1 Normalized Graph (Interactive)

Click below to open the interactive graph in a new window:

Final Summary

The normalized DF1 dataset:

• Integrates complex original structures into a unified schema
• Optimized for multimodal AI and knowledge graph learning
• Includes reproducibility and quality validation
• Supports multiple benchmarks in a single data model

➡️ Core input for FEL‑based clothing recognition, retrieval, and recommendation systems.