SpectralWaste: Multimodal Segmentation
in Waste Management Plants
This project is a collaboration between the RoPeRT research lab from the University of Zaragoza and ATRIA Innovation.
Participants: Sara Casao, Fernando Peña, Alberto Sabater, Rosa Castillón, Darío Suárez, Eduardo Montijano, Ana C. Murillo.
Brief Description
In this project, we use data from a real waste processing facility specialized in plastics, cartons, and cans. The data was collected with a true-to-life prototype of the conveyor belt installed on the waste separation line, closely mimicking the actual installation. This ensures that the waste streams captured accurately mirror those arriving at the facility for separation. Two synchronized cameras were installed for image gathering: a line-scan RGB camera (Teledyne DALSA Linea) and line-scan hyperspectral sensor (Specim FX17) that captures 224 contiguous spectral bands in a range from 900 to 1700 nm.
In the collected images, the objects for automatic identification were selected based on the requirements of the facility. Each class represents elements that commonly cause operational problems in recycling lines, impacting the efficiency of the sorting process. Among these problems, machinery jams pose a significant issue, causing a complete stoppage of the process until the obstructing object is removed. These objects include film and basket, large objects that can clog the conveyor belts as they are not easily breakable; video tape and filament, representing long objects prone to entangling with mechanical parts and requiring manual intervention; trash bag, which encompasses closed bags containing waste that need to be mechanically opened for further processing; and cardboard, paper objects whose recovery adds value by sending them to another recycling process.
SpectralWaste Dataset
Publication: Casao, S., Peña, F., Sabater, A., Castillón, R., Suárez, D., Montijano, E., Murillo, A. C. (2024). SpectralWaste Dataset: Multimodal Data for Waste Sorting Automation. arXiv:2403.18033.
Data and annotations: The set of multimodal images (RGB and HSI) annotated for semantic segmentation that were used in the experiments can be found here.
Auxiliary code and models: Coming soon.
For access to the complete dataset (labeled and unlabeled data) please contact the authors.