Sneaker Authenticity Classifier Demo by Venkata Anirudh ParakalaSneaker Authenticity Classifier Demo by Venkata Anirudh Parakala

Sneaker Authenticity Classifier Demo

Venkata Anirudh Parakala

AI Engineer

Streamlit

Fashion & Apparel

🧠 Sneaker Authenticity Classifier (Demo Stub)

This project is a demo version of a sneaker image classification and authenticity verification pipeline, built with deep learning and Streamlit.

⚠️ Note: This repository contains only stubbed frontend logic and visual samples for demonstration purposes.

🧩 Project Overview

The goal of this project is to build a 2-stage AI system that can:

Identify the sneaker model (e.g., Jordan 1, Yeezy 500, Off-White AJ1)

Verify authenticity (real vs fake) using model-specific visual patterns such as:

stitching

toe box shape

insole branding

label formatting

✅ A human-in-the-loop review step enables semi-automated decision making with Grad-CAM heatmaps for transparency.

🔍 Key Features

✅ 2-Stage Architecture

Stage 1: Model classification using EfficientNet-B0

Stage 2: Context-aware authenticity classifier

Uses predicted model label as auxiliary input

✅ Grad-CAM heatmaps Generated for all "fake" predictions to assist human reviewers

✅ Human-in-the-loop Review A backend reviewer confirms or overrides AI predictions before final verdict

✅ Streamlit Interface Clean frontend UI for uploading images and viewing results

📈 Model Performance

Model Classifier (`model_classifier_final.pth`)

Phase Accuracy Frozen (3 epochs) 71.48% Fine-tuned (12 epochs) 99.61% ✅

Authenticity Classifier (`context_auth_classifier_final.pth`)

Phase Accuracy Frozen (3 epochs) 71.48% Fine-tuned (12 epochs) 94.34% ✅

Combined End-to-End Accuracy:

Estimated pipeline accuracy: ~93–95% depending on image quality

🧠 Technical Highlights

🔁 Transfer Learning via EfficientNet-B0 (pretrained on ImageNet)

🧩 Multi-input model for authenticity (image + model label one-hot)

🎨 Explainability: Grad-CAM visualizations

🧪 Manual curation for verified authentic and fake samples from trusted datasets (e.g., GOAT, StockX, Reddit replicas)

📷 Sample Visuals

Prediction Grad-CAM Overlay Yeezy 500

UI Screenshot

🚀 Try the App Locally (Stubbed Version)

This repo only runs a stub version of the interface:

pip install -r requirements.txt
streamlit run streamlit_ui_stub.py

⭐️ Future Scope

Expand support to 20+ sneaker models including limited editions

Convert the project to a mobile app using PyTorch Mobile or TensorFlow Lite

Build an admin dashboard for human reviewers to accept/reject predictions

Integrate resale marketplace APIs (e.g., StockX, GOAT) for instant comparisons

Offer this as a SaaS platform or plugin for sneaker marketplaces.

🧑‍💼 Author & Developer

Developed by Venkata Anirudh Parakala 🎓 M.S. Computer Science | Sneaker enthusiast + AI developer 📧 Contact: anirudhparakala@gmail.com 🔗 LinkedIn: linkedin.com/in/anirudh-parakala

⚠️ Disclaimer

This is a public-safe wrapper of a proprietary deep learning pipeline. The model training logic, datasets, and weights are not included in this repo.

Like this project

Posted Aug 10, 2025

Created a sneaker authenticity classifier demo using deep learning and Streamlit.

Likes

Views