AI-Powered Visual Search

Find any shoe in seconds, not shelves

Photograph a shoe and instantly get its matching catalog design, confidence score, and exact shelf location. No more manual searching through inventory — just point, shoot, and locate.

Launch Match Studio
ShoeMatch AI Catalog Sample SHOE-001
5-Stage Visual Search Pipeline

From Camera Frame to Warehouse Slot

Our deep learning vision architecture processes raw floor photos through five robust stages to retrieve exact spatial location coordinates.

01

Camera Photo

Worker snaps a shoe photo on warehouse floor or camera upload.

02

U2-Netp Cutout

Saliency neural net strips background clutter & isolates shoe shape.

03

DINOv2 Feature

Vision Transformer extracts a 384-dim dense feature embedding.

04

Footwear Gate

Linear classifier gates non-shoe items & slipper submissions.

05

FAISS Slot Match

Cosine vector search returns ranked SKU and exact shelf location.

Performance Architecture

Dense Vector Search vs. Classical Histograms

Why classical color histograms fail under warehouse lighting variations — and how DINOv2 self-supervised embeddings deliver robust matching accuracy.

~38ms Vector Query Latency

Powered by FAISS `IndexFlatIP` inner-product cosine similarity search operating over 384-dimensional floating point embeddings.

EXIF & Rotation Invariance

Camera uploads undergo automatic EXIF orientation baking (`ImageOps.exif_transpose()`) to maintain accuracy across physical phone orientations.

Zero Slipper False-Positives

Dual-signal footwear classifier head guarantees slippers and non-shoe objects are filtered out cleanly (`reason='slipper_rejected'`).

Warehouse & Logistics

Eliminating the Manual Shelf Search

Compare traditional unlabelled box hunting with point-and-shoot visual retrieval.

Traditional Inventory Search
  • • 15 to 25 minutes spent walking warehouse aisles per sample match
  • • Opening dozens of unlabelled footwear boxes manually
  • • High human error from mislabeled box tags or lost archive samples
ShoeMatch AI Instant Retrieval
  • • Sub-second camera photo lookup directly on phone or tablet
  • • Exact storage coordinate output: Building → Section → Rack → Shelf
  • • Full product metadata attached (upper/sole materials, season, style)
Multi-Platform Ready

Built for Factory & Warehouse Operations

Available both as a responsive web app and a native Android app (`in.co.aflix.shoematchai`) built with Capacitor 6.

Native Camera Integration

Direct hardware shutter integration via `@capacitor/camera` for one-handed photo capture on the warehouse floor.

Glove-Friendly Touch Targets

Material Design 3 minimum 48×48dp controls designed specifically for workers using protective factory gloves.

Role-Based Access Control

JWT authentication supporting distinct `employee` search views and `admin` management panels with audit logging.

Open Source SOTA Architecture

Tech Under the Hood

We leverage state-of-the-art vision models and high-speed vector indexing.

DINOv2 (Meta AI)

Self-Supervised Vision Transformer

Extracts 384-dimensional dense visual embeddings invariant to lighting changes.

U2-Netp Neural Net

Saliency Object Segmentation

Lightweight U-Net isolates shoe foreground cutouts from complex factory clutter.

FAISS IndexFlatIP

Inner-Product Vector Search

Facebook AI Similarity Search engine executes sub-second cosine distance ranking.

FAQ

Frequently Asked Questions

Our pipeline includes a dedicated U2-Netp saliency segmentation step that isolates the shoe foreground cutout before computing visual embeddings. This ensures background clutter, warehouse floor tiles, or hands holding the sample do not disrupt match accuracy.
ShoeMatch AI features a dual-signal footwear classifier head. Submissions classified as non-footwear or slippers are gated cleanly with a friendly `slipper_rejected` alert state, preventing invalid catalog queries.
Yes! The project is packaged both as a web app (accessible via browser) and a native Capacitor 6 Android app (`in.co.aflix.shoematchai`) supporting live hardware camera capture and offline JWT token caching.
Admins can upload multi-angle photos (side, top, front, sole, detail) along with product metadata (materials, season) and exact storage coordinates (Zone → Shelf → Drawer → Slot). DINOv2 feature vectors are automatically generated and appended to the FAISS index.

Eliminate Warehouse Search Latency Today

Deploy ShoeMatch AI across your factory floor or test visual matching in your browser.

Launch Match Studio Now