[Paper Review] Waste detection in Pomerania: non-profit project for detecting waste in environment

Sylwia Majchrowska, Agnieszka Mikołajczyk+5 more|arXiv (Cornell University)|May 12, 2021

Municipal Solid Waste Management56 references4 citations

TL;DR

This paper presents an open-source, two-stage deep learning framework for detecting and classifying litter in diverse environments, using a detector network for localization and a classifier for seven waste categories. It achieves 70% average precision in detection and 75% classification accuracy, offering a scalable solution for automated waste management and environmental monitoring using publicly available datasets and models.

ABSTRACT

Waste pollution is one of the most significant environmental issues in the modern world. The importance of recycling is well known, either for economic or ecological reasons, and the industry demands high efficiency. Our team conducted comprehensive research on Artificial Intelligence usage in waste detection and classification to fight the world's waste pollution problem. As a result an open-source framework that enables the detection and classification of litter was developed. The final pipeline consists of two neural networks: one that detects litter and a second responsible for litter classification. Waste is classified into seven categories: bio, glass, metal and plastic, non-recyclable, other, paper and unknown. Our approach achieves up to 70% of average precision in waste detection and around 75% of classification accuracy on the test dataset. The code used in the studies is publicly available online.

Motivation & Objective

• To address the challenge of unstructured, real-world litter detection and classification using deep learning.
• To develop a unified, open-source system capable of detecting and categorizing waste in varied environmental conditions.
• To improve recycling efficiency by automating waste classification based on real-world data and local sorting rules from Gdańsk, Poland.
• To create a comprehensive, publicly available dataset and model framework for researchers and practitioners in environmental AI.
• To enable applications such as drone-based illegal dump detection and smart bin automation through robust object detection and classification.

Proposed method

• A two-stage deep learning pipeline: first, a detector (EfficientDet-D2) localizes litter in images using anchor-based object detection.
• Second, a classifier (EfficientNet-B2) assigns one of seven categories: bio, glass, metal and plastic, non-recyclable, paper, other, or unknown.
• The dataset combines multiple public sources, including urban, natural, and underwater waste imagery, with careful annotation and data augmentation.
• Pseudo-labeling was applied to semi-labeled data (OpenLitterMap) to improve classification performance, though gains were limited.
• The framework uses transfer learning and hyperparameter tuning to optimize detection and classification on real-world, uncontrolled scenarios.
• The system is fully open-source, with code and trained models publicly available on GitHub.

Experimental results

Research questions

• RQ1 Can a deep learning-based system achieve high-precision detection and classification of litter in uncontrolled, real-world environments?
• RQ2 How effective is a two-stage detector-classifier pipeline for classifying diverse waste types across urban, natural, and aquatic settings?
• RQ3 To what extent does data augmentation and pseudo-labeling improve classification accuracy on imbalanced and partially labeled waste datasets?
• RQ4 How does the proposed framework compare to existing methods in terms of detection mAP and classification accuracy on public benchmarks?
• RQ5 Can the system be generalized to support applications like drone-based illegal dump detection or automated waste sorting in smart bins?

Key findings

• The framework achieved 66.4% average precision (AP@0.5) in litter detection using the EfficientDet-D2 model, outperforming prior methods on comparable datasets.
• The classification model reached 75% accuracy on the test set, demonstrating strong performance despite class imbalance and limited labeled data.
• The use of pseudo-labeling provided only a marginal improvement, suggesting limitations in data quality and imbalance for rare classes like bio or other waste.
• The system successfully localized medium and large litter objects with high precision but struggled with small debris, indicating a need for improved small-object detection.
• The open-source framework and dataset are publicly available, enabling reuse for environmental monitoring, smart waste management, and further research.
• The results demonstrate the feasibility of deploying deep learning for real-world waste detection, with potential for integration into drones or robotic sorting systems.