Garbage-Image-Classification

Real-Time Waste Classification using MobileNetV3

This repository contains my contributions to the Garbage-Crew project built during the 2025 UTD Deep Dive AI Summer Workshop. It focuses on the image classification component of an AI-powered smart waste sorter.

---

What This Repo Includes

• Dataset Preparation: Preprocessing and synthesis of a unified waste classification dataset from multiple public sources.
• Model Training: Fine-tuned a pretrained MobileNetV3 model on 8 waste categories using PyTorch and timm.
• Real-Time Classification: Python script using OpenCV and the trained model to classify waste items in a live webcam feed.

---

Dataset

Unified across multiple sources and organized into 8 classes:

• battery, glass, metal, organic_waste, paper_cardboard, plastic, textiles, trash

Available on Kaggle: Unified Waste Classification Dataset

---

Model

• Architecture: mobilenetv3_large_100 from timm
• Input Size: 224x224
• Framework: PyTorch
• Loss: CrossEntropyLoss
• Optimizer: Adam
• Training: 80/20 stratified split with basic preprocessing (resize, normalize)

---

Live Webcam Classifier

• Uses your system’s webcam to classify waste in real-time
• Optional object detection feature using YOLOv8 for bounding boxes
• Snapshot mode: Press 's' to save the current frame with classification results

---

Requirements

• Python 3.9+
• torch, timm, opencv-python, torchvision, numpy

Install dependencies:

pip install -r requirements.txt

---

Full Project

This repository focuses only on data + model + real-time classification.

For the complete system including hardware integration (EV3 motor, Raspberry Pi, physical trapdoor mechanism), visit:

Garbage-Crew/Garbage-Crew