AI for resource-efficient circular fashion
| Reference number | |
| Coordinator | Wargön Innovation AB |
| Funding from Vinnova | SEK 7 000 000 |
| Project duration | October 2021 - April 2024 |
| Status | Completed |
| Venture | AI - Leading and innovation |
| Call | AI in the service of the climate 2 |
Important results from the project
The main objective was to identify, design and develop AI-solutions for textile sorting, with an explicit focus on automation of reuse sorting processes. The overall purpose was to enable a sizable increase of the utilization rate per garment, thus offsetting the need for new production which enables environmental gains. . The two overall objectives were: 1. Creation of an open dataset for second hand garments, with at least 30 000 garments (photos and information). 2. AI model prototypes for potential AI user needs, evaluated by “AI vs human” pilots. These objectives were fulfilled.
Expected long term effects
An open dataset with 31,997 objects has been published including three photos per garment (front, back and brand) and associated information (annotations) such as garment type, colour, material and condition. Prototypes of AI models to assess the various annotations have been developed and tested. An LCA has been done for the process and it shows that by achieving as little as 1% increased reliability in sorting, AI models have a positive environmental impact. Expected effect is that different stakeholders use the data in the dataset for development of AI models for textile sorting.
Approach and implementation
The work process for development of the dataset and prototype was needs-driven via RISE UX-team, who mapped the needs of all project partners. An AI-annotationstation with cameras and software was initially built at Wargön Innovation so that data collection could start and improve. Later, a station at Myrorna was also added. At the same time, RISE AI-team developed AI-models from the information in the dataset. The data collection was completed late in the project and affected how well the AI models could be trained due to insufficient amount of data.