Comprehensive Comparison of Major Fashion Resources

1) Side-by-side Comparison

Category	DeepFashion	DeepFashion2	DeepFashion3D	MMFashion	Fashionpedia
Resource Type	Dataset + Model	Dataset + Benchmark	Dataset + Methodology	Platform / Toolkit	Ontology + Dataset
Year / Venue	2016 / CVPR	2019 / CVPR	2020 / CVPR	2020 / arXiv	2020 / CVPR
Primary Goal	Large-scale 2D clothing recognition & retrieval	Real-world, multi-task fashion understanding	Single-image 3D garment reconstruction	Unified execution framework for fashion vision tasks	Structuring fashion knowledge via explicit ontology
Problem Motivation	Limited scale & labels in earlier datasets	Sparse landmarks & single-garment bias in DF	Severe lack of 3D garment data	Fragmented fashion AI codebases	Inconsistent definition of fashion concepts
Data Dimensionality	2D images	2D multi-instance scenes	3D meshes + images	2D (uses external datasets)	2D pixel-level annotations
Dataset Scale	800K+ images	491K images / 801K garments	2,078 3D garment models	No proprietary dataset	48K+ images
Garment Categories	50+	13	~10 garment types	Dataset-dependent	27
Core Annotations	Category, attributes, sparse landmarks	BBox, mask, dense landmarks (39), pose, pairing	Mesh, UV, camera, pose, feature lines	Models, configs, pipelines	Instance, part, attribute, mask
Landmark Concept	Sparse 2D keypoints	Dense pose-aware landmarks	3D structural feature lines	Task-dependent module	Encoded via part boundaries
Pose Information	❌	Clothing pose	Body–garment pose (SMPL)	Task-dependent	❌
Segmentation	❌	Instance mask	Implicit via mesh surface	Supported	Part-level masks
Attributes	Image-level	Image-level	Texture (not semantic)	Supported	Part-localized
Ontology Explicitness	❌	❌	❌	❌	✅ Core contribution
Main Tasks	Classification, retrieval, attribute prediction	Detection, pose, segmentation, Re-ID	3D reconstruction, registration, texture recovery	Attribute, retrieval, parsing, compatibility	Detection, segmentation, attribute localization
Proposed Model	FashionNet	Match R-CNN baseline	Hybrid mesh + implicit surface	Backbone–Head modular framework	Mask R-CNN baseline
Methodological Core	Landmark-aware pooling	Multi-task unified evaluation	Template adaptation + physics-aware features	Modular engineering design	Ontology-driven annotation
Evaluation Metrics	Top-k accuracy, recall	AP, OKS, PCK, ReID recall	Chamfer, EMD, Normal Consistency, 3D IoU	Task-standard metrics	mIoU, AP (det/attr)
Major Strength	First large-scale fashion dataset	Real-world complexity & multi-task scope	Provides true 3D structural evidence	Easy experimentation & extensibility	Explicit semantic structure of fashion
Main Limitation	Lacks structural & semantic depth	Limited semantic/ontological reasoning	Computationally heavy, frontal bias	Limited theoretical contribution	No aesthetic/emotional modeling
Emotion / Aesthetics Modeled	❌	❌	❌	❌	❌

Tip: On smaller screens, scroll horizontally to view all columns.

2) Conceptual Evolution in Fashion Vision

Stage	Resource	Key Question Answered
Appearance Level	DeepFashion	What clothing is visible?
Structural 2D Level	DeepFashion2	Where are parts under real-world variations?
Semantic-Part Level	Fashionpedia	Which part has which property?
Geometric 3D Level	DeepFashion3D	What is the physical garment shape?
Engineering Layer	MMFashion	How do we implement and experiment efficiently?

3) Research-Oriented Selection Guide

Research Goal	Most Suitable Resource
Attribute learning	DeepFashion, Fashionpedia
Detection in complex scenes	DeepFashion2, Fashionpedia
Landmark / garment pose	DeepFashion2
Part-based reasoning	Fashionpedia
Clothing retrieval / Re-ID	DeepFashion, DeepFashion2
Virtual try-on / simulation	DeepFashion3D
3D garment reconstruction	DeepFashion3D
Multi-task unified training	Fashionpedia + DeepFashion2
Rapid experimentation	MMFashion

4) Big-Picture Insight

These resources form complementary layers rather than replacements: DeepFashion → DeepFashion2 → Fashionpedia → DeepFashion3D progressively moves fashion AI from visual appearance → structural layout → semantic understanding → physical geometry, while MMFashion provides the engineering backbone to operationalize all of them.