MMoRWaD: A Multimodal Dataset for Recyclable Waste Recognition Acquired with an Edge-Node Reverse Vending Machine Station
TL;DR
This work developed a custom edge-node reverse-vending-machine data-collection station equipped with multimodal sensors to co-record RGB-D video streams and impact audio on a single local compute node within the same acquisition window to support research on multimodal waste recognition, sensor fusion, and resource-efficient classification.
Abstract
Automated classification of municipal solid waste remains challenging due to high variability in object appearance, deformation, contamination, and inconsistent consumer disposal behavior, creating a need for representative multimodal datasets that enable robust machine-learning models. In this work, we developed a custom edge-node reverse-vending-machine (RVM) data-collection station equipped with multimodal sensors (RGB-D and microphone array) to co-record RGB-D video streams and impact audio on a single local compute node within the same acquisition window. Strict data acquisition protocols were designed to ensure a consistent environment, controlled illumination, and uniform background conditions while capturing realistic variations associated with everyday Greek consumer waste. The resulting dataset comprises 3300 multimodal samples corresponding to 721 distinct household waste items. Each sample is fully annotated across six categories addressing features critical for recycling: Material (18 classes), Hue (16 classes), Top/Cap, Label, Contamination, and Deformation. In addition, Common Objects in Context (COCO) annotations were added to support segmentation and object detection problems. This dataset is intended to support research on multimodal waste recognition, sensor fusion, and resource-efficient classification, and to serve as a foundation for evaluating models targeting edge-deployed recycling systems.