PET Perplexity ♻️

Team internship_needed (AG41)

PET Perplexity is an intelligent, automated polymer segregation system designed to revolutionize plastic waste management. It uses a cross-platform mobile application powered by advanced Computer Vision and Machine Learning to detect, classify, and analyze PET bottles in real-time.

The system addresses the challenge of segregating plastic waste by identifying key attributes such as bottle presence, size, brand, and material properties.

🚀 Key Features

• Real-time Object Detection — Instantly detects PET bottles within a video feed or captured image.
• Bottle Size Classification — Automatically categorizes bottles into standard sizes (small, medium, large) to aid sorting logistics.
• Brand Recognition — Identifies the brand of a bottle using custom-trained deep learning models.
• Material Analysis — Uses XGBoost to analyze polymer characteristics for precise segregation.
• Batch Scanning — Processes multiple items at once for high-throughput environments.
• User Dashboard — A comprehensive mobile interface for tracking scan history and segregation statistics.

📱 App Preview

Real-time Detection	Size Classification	Brand Recognition
Material Analysis	Batch Scanning

🛠️ Tech Stack

Frontend (Mobile App)

• Framework: Flutter (Dart)
• Platforms: Android, iOS, Web
• State Management: Provider / Riverpod (inferred)
• Architecture: Feature-first architecture (lib/features/)

Backend & Machine Learning

• Languages: Python, Node.js
• Frameworks: Flask, FastAPI / Uvicorn, Gradio
• Computer Vision:
  • YOLOv5 / YOLOv8 — robust object detection and bounding box regression.
  • OpenCV (cv2) — image preprocessing and frame manipulation.
• Deep Learning Models:
  • EfficientNet-B2 — fine-tuned for high-accuracy bottle size classification.
  • Custom CNNs (PyTorch) — brand logo detection and classification.
• Machine Learning:
  • XGBoost — tabular data analysis and material property prediction.

Branches & Deployments

This repo is split across branches by service. main holds the mobile app and two lightweight predictor services; the rest of the pipeline and model training code live in separate branches.

Branch	Contents	Live deployment
main	Flutter app (Android/iOS) + brand_predictor + dim_predictor	NOT CURRENTLY DEPLOYED
ML	Integrated YOLOv5 + CNN + XGBoost pipeline	https://huggingface.co/spaces/SudoKuder/agglo
bottlesize	EfficientNet-B2 bottle-size classifier training + Flask API	NOT CURRENTLY DEPLOYED
backend	Node/Express REST API (auth, scans, MongoDB, Cloudinary)	https://pet-perplexity.onrender.com

To run any service locally: git checkout <branch>, then follow that branch's own README/requirements.txt.

---

Project Structure

Agglomeration-2.0/            (main branch)
├── lib/                      # Flutter app source
├── android/ ios/ web/        # Platform targets
├── brand_predictor/          # Logo detection + brand classification (Gradio, HF Space)
│   ├── app.py
│   ├── pipeline.py           # Saliency transformer + two-stream EfficientNet-B2
│   ├── model_arch.py
│   └── utils.py
├── dim_predictor/             # Dimension estimation (Gradio, HF Space)
│   ├── app.py                # YOLOv8-seg + ArUco/cap-reference/camera-geometry calibration
│   └── README.md
├── Problem Statement.pdf
└── pubspec.yaml
 
Other branches:
├── ML/                        # Integrated YOLOv5 + CNN + XGBoost inference pipeline
├── bottlesize/                 # EfficientNet-B2 training + standalone Flask API
└── backend/                    # Node/Express API + MongoDB + Cloudinary

---

ML Pipeline Architecture — Dimension Layer

Current text says the Dimension Layer uses EfficientNet-B2. What's actually deployed in dim_predictor is:

Dimension Layer: YOLOv8 segmentation isolates the bottle from the background, then physical height/diameter are computed from the mask using a three-way scale-calibration fallback: an ArUco marker if one is visible in frame, otherwise a known bottle-cap diameter as a size reference, otherwise camera geometry/distance. (EfficientNet-B2 was an earlier size-classification approach, now developed separately in the bottlesize branch as discrete volume classes — 33cl/50cl/100cl/150cl/200cl — rather than continuous measurements.)

🤝 Contributing

Contributions are always welcome!

1. Fork the repository.
2. Create a feature branch (git checkout -b feature/AmazingFeature).
3. Commit your changes.
4. Open a Pull Request.

---

👥 Development Team

Name	Role
Aditya	Mobile App & UI/UX Development
Samarth Agarwal	Backend Development
Apurva Arya	AI & ML/DL Model Development
Arnav Tripathi	AI & ML/DL Model Development
Suryansh Kulshreshtha	AI & ML/DL Model Development