Advancing AI-Driven Waste Recovery through Open-Vocabulary Detection, Pseudo-Labeling, and Semi-Supervised Learning
Abid, Hassan
Abid, Hassan
Author
Supervisor
Department
Machine Learning
Embargo End Date
2027-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
The rapid growth of municipal solid waste due to urbanization and increased consumption has escalated the need for efficient waste management. Material Recovery Facilities (MRFs) play a crucial role in sorting recyclable materials, yet their reliance on manual or semi-automated processes leads to operational inefficiencies and exposes workers to significant health risks, including injuries from sharp objects and exposure to toxic substances. Despite the critical importance of waste sorting, the waste domain, especially complex real-world datasets in the context of object detection, remains largely underexplored in current research. In this thesis, we address these challenges by systematically evaluating state-of-the-art open-vocabulary object detection (OVOD) models on the ZeroWaste dataset, a highresolution collection of images captured from a real-world MRF that encapsulates the complexities of cluttered, deformable, and overlapping waste objects. Our zero-shot evaluations, utilizing both basic class-only and enhanced textual queries, reveal that OVOD methods are severely limited by domain gaps, high intra-class variability, and the inherent clutter of industrial waste streams. To address these limitations, we pivot from zero-shot detection to targeted domain adaptation. First, we establish new baselines by fine-tuning state-of-the-art detectors, including recent real-time and transformer-based models such as YOLO11 and DINO, on the labeled ZeroWaste-f subset—nearly doubling the performance of earlier CNN-based baselines. Building on this, we propose an effective semi-supervised learning pipeline that generates high-quality pseudo-annotations for the unlabeled ZeroWaste-s data through an ensemble-based consensus pseudo-labeling strategy combined with soft labeling. This new method effectively reduces label noise, balances prediction uncertainty, and improves overall detection accuracy. Overall, our work provides a comprehensive analysis of current OVOD methods in the challenging waste sorting domain, while establishing new baselines through targeted fine-tuning of advanced detectors and generating high-quality pseudo-annotations for the unlabeled subset using our ensemble-based pseudo-labeling pipeline. Our contributions advance open-source research in waste management by rigorously benchmarking advanced OVOD and closed-set detectors and addressing data scarcity with a publicly released high quality annotated dataset. Together, our results establish new performance baselines and provide critical resources that pave the way for effective AI-driven waste sorting solutions to support sustainable recycling practices.
Citation
Hassan Abid, “Advancing AI-Driven Waste Recovery through Open-Vocabulary Detection, Pseudo-Labeling, and Semi-Supervised Learning,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Waste Management, Object Detection, Open Vocabulary Object Detection, Semi-supervised Learning, Ensemble-based Learning, Pseudo Labeling