BinGo: Real-Time Multi-Item Waste Sorting App

A object detection app for real-time waste sorting, featuring multi-label classification with coarse-material prediction fallback for robust handling of novel objects.

Demo

Highlights

• Multi-Item Detection: Identifies multiple waste items simultaneously using YOLOv8
• Hierarchical Classification: Two-level taxonomy (5 materials → 14 objects) with fallback mechanism
• 114% mAP Improvement: From 0.36 (baseline) to 0.77 through dataset optimization
• Mobile Ready: TensorFlow Lite deployment for real-time Android inference

Workflow

                    ┌─────────────────┐
                    │   5 Materials   │
                    │ Metal│Glass│... │
                    └────────┬────────┘
                             │ fallback when uncertain
┌──────────┐    ┌───────┐    ▼
│  Image   │───>│ YOLO  │───> 14 Object Types ───> Regional Bin Mapping
└──────────┘    └───────┘    PET│HDPE│Wrapper...    (Garbage/Recycle/Mixed)

Quick Start

Installation

git clone https://github.com/PFCS33/waste-sort-yolo.git
cd waste-sort-yolo

# Create environment
conda create -n bingo python=3.10 -y
conda activate bingo

# Install dependencies
pip install -r requirements.txt

Inference

python main.py --source path/to/image.jpg --weights weights/best.pt

Dataset

We constructed a balanced 14-class dataset by merging 13+ public sources.

Metric	Baseline Dataset	Our Dataset
Classes	7 (materials)	14 (objects)
Imbalance Ratio	>50	4.84
Images	~5k	~37k

Class Taxonomy

Material	Object Types
Plastic	PET Container, HDPE Container, Plastic Wrapper, Plastic Bag, Plastic Cup, Styrofoam, Tetra Pak
Paper	Paper, Paper Cup, Paper Bag, Used Tissue
Metal	Metal
Glass	Glass
Cardboard	Cardboard

Data Preparation

Our automated pipeline downloads, transforms, and merges multiple public datasets using a YAML configuration.

YAML Configuration Template:

global:
  root_dir: "data_merge"          # Output directory
  target_format: "yolo"           # Annotation format
  nc: 14                          # Number of classes
  split: [0.8, 0.1, 0.1]          # Train/val/test split ratio
  classes:                        # Unified class names
    - METAL
    - GLASS
    - PET_CONTAINER
    # ... (14 classes total)

datasets:
  - name: "dataset-name"
    download:
      source: "roboflow"          # Source: roboflow | kaggle
      params: ["workspace", "project", version, "yolov8"]
    transform:
      exclude_class: [0, 2]       # Classes to exclude (optional)
      class_mapping:              # Map source labels to unified labels
        0: 8                      # source_id: target_id
        1: 1
    paths:                        # Image/label paths in source dataset
      - images: ["train", "images"]
        labels: ["train", "labels"]

Run the pipeline:

bash run_data.sh

Model Zoo

Model	mAP@0.5	mAP@0.5:0.95	Params	Download
Baseline (7-class)	0.54	0.36	3.2M	-
Improved (14-class)	0.92	0.77	3.2M	weights
Hierarchical (BCE)	0.89	0.75	3.2M	weights
Hierarchical (BCE+penalty)	0.90	0.76	3.2M	weights

Training

# Standard training
python main.py --mode train --data configs/data.yaml --epochs 300

# Hierarchical multi-label training
python main.py --mode train --loss bce_penalty --epochs 300

Results

Performance Comparison

Stage	Approach	mAP@0.5:0.95	Novel Object Acc
1	Baseline (7-class)	0.36	23%
2	Fine-grained (14-class)	0.77	-
3	Hierarchical + Fallback	0.76	64%

Project Structure

waste-sort-yolo/
├── main.py                 # Entry point for training/inference
├── run_data.sh             # Dataset preparation script
├── weights/                # Pretrained model weights
└── scripts/
    ├── data/               # Dataset pipeline
    │   ├── config.yaml     # Dataset configuration
    │   ├── config_h.yaml   # Hierarchical config
    │   └── utils/          # Download, transform, merge utilities
    ├── models/
    │   ├── baseline/       # Standard YOLOv8 training
    │   └── multi_label/    # Hierarchical multi-label YOLO
    │       ├── config.py   # Hierarchy configuration
    │       ├── dataset.py  # Multi-hot label generation
    │       ├── loss.py     # BCE & hierarchical penalty loss
    │       ├── nms.py      # NMS with fallback mechanism
    │       ├── predictor.py
    │       └── trainer.py
    └── utils.py

Related Repositories

Repository	Description
panqier/vcproject	Android app with TFLite deployment
ultralytics/ultralytics	YOLOv8 base implementation

Acknowledgments

• Ultralytics for YOLOv8 implementation
• TACO, TrashNet, and other public datasets
• Course project for CMPT 742 Visual Computing at Simon Fraser University

License

This project is licensed under the MIT License - see the LICENSE file for details.