AI powered Object Detection
eCommerce App

eCommerce App

The Challenge

Fashion items are complex to identify due to variations in style, color, texture, size and background noise. Traditional image search systems struggle with:

  • Accurate detection of multiple fashion items in a single image
  • High-precision similarity matching across diverse catalogs
  • Real-time performance on mobile platforms

Solution

Advanced Object Detection

  • Implemented YOLO + DINO models to detect and isolate fashion items such as tops, trousers, bags, and accessories
  • Achieved ~90% detection accuracy across real-world images with complex backgrounds
  • Enabled precise cropping and normalization to improve downstream similarity search
eCommerce App Solution
Circle

Key Features

AI-Based Fashion Object Detection

The platform uses a production-grade OWL-ViT based detection pipeline enhanced with a custom multi-stage filtering architecture.

Supported Detection Categories
The system detects multiple fashion categories including:
Shirts, T-Shirts, Blouses, Jackets, Blazers, Pants, Jeans, Dresses, Shoes, Bags, Hats, Jewelry, Sunglasses

The platform also supports simultaneous multi-object detection within a single image.

Three-Gate Detection Optimization Pipeline

To improve detection quality and reduce false positives, AppleTech engineered a custom three-
stage detection pipeline.

Gate 1 — Label-Specific Thresholding

Each fashion category uses custom confidence thresholds.

Examples:

  • Rings require stricter confidence
  • Tops allow broader recall

This eliminated low-confidence noisy detections early in the pipeline.

Gate 2 — Layer-Aware Filtering

Different filtering rules were applied for:

Examples:

  • Garments
  • Accessories
  • Jewelry

This improved precision across fashion layers while maintaining recall.

Gate 3 — Semantic Validation

Final semantic filtering ensured:

Examples:

  • Stable detections
  • Cleaner crops
  • Better downstream retrieval quality

This significantly improved similarity-search performance.

Intelligent Bounding Box Optimization

One of the major challenges was inaccurate torso-level detections caused by face and head
overlap.

AppleTech introduced:

  • IOU-based overlap suppression
  • Face-aware bounding-box corrections
  • Garment hierarchy rules
  • Dress vs top-bottom conflict resolution

These enhancements improved crop quality and significantly boosted retrieval accuracy.

Fashion Similarity Search Engine

The platform includes a highly optimized fashion similarity retrieval engine.

AI Embedding Pipeline

The workflow includes:

1. Detect fashion item
2. Crop detected region
3. Generate FashionCLIP embeddings
4. Search embeddings in Weaviate
5. Re-rank results using MMR (Maximum Marginal Relevancetu)

This enabled visually relevant and diversified product recommendations.

Vector Database & Semantic Search

The system leveraged:

  • Weaviate for vector similarity search
  • FashionCLIP embeddings for semantic understanding
  • PostgreSQL for metadata storage

Both image embeddings and text embeddings were indexed for:

  • Visual search
  • Semantic product search
  • AI chatbot retrieval
  • Outfit matching

The indexing strategy reduced duplicate embeddings and improved retrieval diversity by
indexing unique product-color combinations instead of every SKU individually.

Automated Fashion Data Pipeline

AppleTech built a scalable fashion scraping and indexing infrastructure.

Features Included

  • Sitemap-based URL discovery
  • Automated product scraping
  • Variant extraction
  • Image normalization
  • Lazy-loading image handling
  • Metadata extraction
  • Offline recovery pipelines

The system supported multiple fashion brands and could continue operating even when third-
party websites changed frontend structures.

Technology

Python
React
PostgreSQL
Weaviate

Results & Benefits

Detection Accuracy

  • Achieved approximately 90% detection accuracy for supported fashion categories
  • Reliable performance across real-world images with complex backgrounds

Similarity Search Improvements

  • High-quality image-to-image retrieval
  • Improved recommendation diversity using MMR reranking
  • Faster semantic fashion search using vector embeddings

Operational Benefits

  • Scalable microservice-based architecture
  • Reduced duplicate retrievals
  • Improved indexing efficiency
  • Faster product discovery experience for end users

Technical Challenges Solved

Complex Fashion Classification

Resolved ambiguity between:

  • Dresses vs tops
  • Jackets vs blazers
  • Shorts vs skirts

Small Object Detection

Improved detection of:

  • Rings
  • Earrings
  • Bracelets
  • Sunglasses

Multi-Garment Understanding

Implemented logic to better handle:

  • Layered clothing
  • One-piece garments
  • Overlapping apparel

Retrieval Diversity

MMR-based reranking reduced near-duplicate recommendations and improved user experience.

Technical Challenge AI App
Circle