DeepFashion3D (DF3 V2) — Normalized Dataset for FEL

1. Dataset Overview

DeepFashion3D is a large-scale benchmark for 3D garment reconstruction from a single RGB image. While DF1/DF2 focus on 2D evidence (bbox/landmarks/segmentation/re-ID), DF3 enables evaluation of true 3D garment geometry including wrinkles, thickness, and asymmetry.

In this pipeline, DF3 outputs (df3_fel_out.zip) are normalized into a graph-friendly, join-stable reference layer for FEL. Instead of copying or fully parsing heavy 3D assets (OBJ/PLY/PKL/PNG), the extractor builds index tables + a pose-centric hub so that downstream loaders can parse 3D contents safely and on-demand.

Key Characteristics

• Primary join key: pose_key = garment_id-pose_id
• Hub table: items.csv.gz (one row per pose_key; modality coverage flags + representative file refs)
• Modality indices separated by type (pose params / pointcloud / mesh bundle / feature lines)
• Provenance + auditability via manifest.* and qc_summary.json
• QC status (artifact basis): hard_fail = False, warnings/errors = 0

➡️ This converts DF3’s file-centric 3D structure into a standardized, joinable schema optimized for knowledge-graph construction and multimodal evidence linkage.

2. Folder and File Structure

(1) Original DF3 Structure (Input)

The following input structure is reconstructed from manifest path patterns and extractor auto-discovery rules:

• featureline_annotation/ → feature line (curve) annotations as .ply
• point_cloud/<garment_id>/<garment-pose>.ply → per-pose point clouds
• packed_pose/<garment_id>/<garment-pose>.pkl → per-pose parameters
• filtered_registered_mesh-001/<garment-pose>/ → mesh bundle (.obj, .mtl, *_tex.png)
• garment_type_list.txt → garment_id → garment_type mapping

(2) Normalized Output Structure

Normalized outputs are written to out_dir/normalized/:

Core Hub + Indices

• items.csv.gz → PoseItem hub (pose_key universe)
• poses_index.csv.gz → pose PKL index
• pointcloud_index.csv.gz → point cloud PLY index (+ optional vertex_count)
• meshes_index.csv.gz → mesh bundle index (OBJ/MTL/TEX refs + existence QC)
• featureline_index.csv.gz → featureline PLY index (1 file = 1 row; includes curve_type_id)
• garments.csv.gz → garment_id aggregation

Management & Validation

• manifest.csv / manifest.jsonl → input audit
• qc_summary.json → run-level QC / coverage / stats

3. Role in FEL

DF3 is pose-centric: the primary semantic unit is a 3D pose instance (pose_key), not an image.

Node Construction

• Garment → garment_id
• PoseItem → item_uid = DF3:pose/<pose_key>
• Evidence refs (indices): PosePKLFileRef / PointCloudFileRef / MeshFileRef / FeatureLineFileRef

Relationship Construction

Garment ─has_pose──────────▶ PoseItem
PoseItem ─has_pose_params──▶ PosePKLFileRef
PoseItem ─has_pointcloud───▶ PointCloudFileRef
PoseItem ─has_mesh─────────▶ MeshFileRef
PoseItem ─has_featureline──▶ FeatureLineFileRef (one-to-many)

Core Design Principles

Single Join Key
All modalities join via pose_key

Evidence Separation
Heavy 3D contents are not parsed here; only references + coverage/QC are normalized

Provenance Preservation
Auditability via manifest.* and run validation via qc_summary.json

4. Extracted Benchmarks

DF3 is not benchmark-split like DF1/DF2; instead it is organized by modalities over the pose_key universe.

• pose_pkls: 1,783 pose_key
• pointclouds: 1,783 pose_key
• meshes: 1,212 pose_key
• featurelines_rows: 7,952 (1 featureline PLY = 1 row)
• pose_keys_union: 1,783 (matches items.csv.gz rows)

Coverage/QC highlights:

• meshes_missing_obj = 0
• meshes_missing_tex = 0
• only_pose_pkl = 0, only_pointcloud = 0, only_mesh = 0

5. Graph Structure Description

Central Node

Items — the pose_key hub that integrates all modalities.

Pose-based Modality Joins

Items ──(pose_key)──▶ PosesIndex
Items ──(pose_key)──▶ PointcloudIndex
Items ──(pose_key)──▶ MeshesIndex
Items ──(pose_key)──▶ FeaturelineIndex

Garment-level Aggregation

Items ──(garment_id)──▶ Garments (aggregation layer, not a strict FK dependency)

Meta Layer

Manifest and QCSummary are run-level metadata (audit/validation), typically shown as dashed conceptual edges.

DF3 Normalized Graph (Interactive)

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

Final Summary

• DF3 is normalized around pose_key to unify multi-modal 3D evidence under one stable join key.
• Index tables provide reference stability, coverage measurement, and safe downstream parsing.
• QC and manifest provide reproducibility and auditability for FEL ingestion.

➡️ Core FEL input for pose-centric 3D garment evidence linking and downstream 3D reconstruction/registration research.