Fashion Product Image Retrieval
Project 1: Attribute-Based Text-to-Image Ranking using CLIP Shared Latent Space
Search Query & Attribute Filter
Top Retrieved Candidates
Loading
Computing CLIP similarity dot product
Ground Truth Filters & Exploration Mode
2,014 relations foundGround Truth Relations List (Query ↔ Product)
Performance Summary by Query Ambiguity Level
| Ambiguity Level | Precision@5 | Precision@10 | Precision@20 | Recall@5 | Recall@10 | Recall@20 |
|---|
Research Queries & Quantitative Evaluation Table
Click the Test button on any query row to evaluate retrieval in real-time
Top 10 Article Type Distribution
Top 10 Base Colours Distribution
Target User Segmentation (Gender)
Product Usage Context Distribution
ArticleType Distribution (13 Categories)
| Article Type | Count | Percentage |
|---|
Top 10 Base Colour Distribution
| Base Colour | Count | Percentage |
|---|
Verified Dataset Sample & Attribute Annotation Table
Quality Audit & Analysis of Image–Metadata Pairs Across 20 Representative Fashion Catalog Samples (Appendix A)
5 Key Multimodal Retrieval Insights for CLIP Zero-Shot Ranking
6 End-to-End Data Preparation Pipeline Stages
Transparansi penuh di seluruh transformasi data: mulai dari verifikasi integritas file fisik mentah, penyaringan kategori target, sampling terstratifikasi yang reproducible, dan sinkronisasi ground truth deterministik, hingga pra-komputasi indeks vektor CLIP ViT-B/32 secara offline.