A Custom YOLO based waste detection system for real-time waste classification through web interface. Classifying waste into 6 categories: Biodegradable, Cardboard, Glass, Metal, Paper, and Plastic.
- Custom YOLO with 14×14 grid detection, 2 boxes per cell, 6 classes
- Flask web application with interactive UI
- Trained on 7,260 images, validated on 3,114 images
- Type: Custom YOLOv1
- Input Size: 448×448 pixels
- Grid Size: 14×14 (S=14)
- Bounding Boxes per Cell: 2 (B=2)
- Classes: 6 waste categories
- Output: Class predictions + bounding box coordinates
Original: Roboflow Garbage Classification
Preprocessed: Custom Kaggle Dataset
- Analyzed dataset with
preprocessing/analyze.py - Filtered images with >98 objects using
preprocessing/max_obj_filter.py - Final: 7,260 images across 6 classes
- Fork notebook: Kaggle Notebook Link
- Use dataset: Custom Kaggle Dataset
- Train on Kaggle GPU
- Download
best_model.pth - Live Demo on Hugging Face Spaces: https://huggingface.co/spaces/lavanyanigam/SortNow
- Clone the Repository
git clone https://github.com/lavanyanigam/SortNow.git
cd SortNow- Install Dependencies
pip install -r requirements.txt- Download Trained Model:
-
Google Drive: Download best_model.pth
-
Place it in the project root so that your folder looks like
SortNow/
├── app.py
├── best_model.pth
├── requirements.txt
├── templates/
└── static/
- Run the Application
python app.py- Open in Browser
http://localhost:5001
- Classify Waste
- Notebook: Training Code
- Framework: PyTorch
- Optimizer: Adam
- Learning rate: 1e-4
- Batch size: 16
- Epochs: 100 (Early stopping at Epoch 64)
- Hardware: Kaggle GPU P100
- Early stopping triggered at epoch 64
- Best model at epoch 54 with validation loss: 176.3113
- Final training loss: 65.5796
- Final validation loss: 177.0820
- These results indicate overfitting, likely due to dataset imbalance and limited model capacity.
- Phase 1 - training on YOLOv11 pre-trained
development_iterations/sortnow_yolov11s_training.ipynb - Phase 2 - Single Object Detection
development_iterations/single-obj-detection.ipynb - Phase 3 - Custom Multi-Object Detection model (https://www.kaggle.com/code/lavanyanigam/yolo-from-scratch)