Skip to content

A real-time edge AI app for waste classification and segmentation using YOLOv5, Roboflow, and TensorFlow Lite, trained on augmented TACO data and deployed to Google Play.

Notifications You must be signed in to change notification settings

VictorChenCA/WasteClassify

Repository files navigation

WasteClassify

WasteClassify is an edge AI mobile application for real-time waste classification, detection, and segmentation. Trained on an augmented TACO dataset via Roboflow using YOLOv5 and ResNet-50, models were converted from PyTorch to TensorFlow Lite (now LiteRT) via ONNX and deployed on Android for fully on-device inference without any cloud dependency.

Features

  • Real-time solid waste classification, detection, and segmentation
  • YOLOv5 for detection and instance segmentation
  • Roboflow-based data augmentation (flip, rotate, brightness, etc.)
  • Edge AI deployment pipeline: PyTorch → ONNX → TensorFlow Lite
  • Android app using TensorFlow Lite runtime (LiteRT) for local inference

Tech Stack

Task Tools & Libraries
Training PyTorch, Roboflow, TACO
Conversion ONNX, TensorFlow Lite
App Development Android Studio, Java
Models ResNet-50 (classification), YOLOv5 (detection + segmentation)

Repository Structure

  • App1/ — Android app code
  • Image_Classification_ResNet50_x_Waste_Cl... — Classification notebook
  • Object_Detection_YOLOv5_x_TACO.ipynb — Detection and segmentation notebook

Results

  • YOLOv5 segmentation mAP@0.5: 0.4744
  • Input image size: 736px, batch size: 7
  • ResNet-50 classification accuracy (3-class): 84.36%

Resources

Acknowledgments

Developed through the Computer Science and Informatics Summer Research Experience (CSIRE) Program at Stony Brook University
Mentored by Dr. Ruwen Qin and Dr. Muhammad Monjurul Karim

About

A real-time edge AI app for waste classification and segmentation using YOLOv5, Roboflow, and TensorFlow Lite, trained on augmented TACO data and deployed to Google Play.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published