Automating second-hand fashion

Note

Update February 2025: For more detailed and up-to-date information, please visit fnauman.com/second-hand-fashion.

Automating Second-Hand Fashion

Textile industry is one of the biggest contributors to global pollution and green house emissions. Some statistics from the Nordic council of ministers report 2023:

• It accounts for up to 10% of global green house gas emssions McKinsey 2020.
• The water use of the textile industry exceeds 215 trillion liters per year.
• In Nordic countries, the annual consumption textiles per capita is between 13.5 and 16 kgs.
• Synthetic fiber production, which are harder to recycle than natural fibers, has increased from less than 20% of the global fiber production to 60% today.
• Less than 1% of textiles are recycled every year.
• Many garments are only used 7-8 times before being discarded. If each garment could be worn twice as much, the emissions from the textile industry would be reduced by nearly half.

These alarming trends and statistics have forced the European Union to propose Extended Producer Responsibility (EPR) for textiles. The EPR is a policy approach that makes the producer responsible for the entire life cycle of the product, including the management of the product after its end-of-life. The EPR is expected to be implemented in the EU by 2025. See here for more details.

AI in the Fashion Industry

AI has been used in many applications in the fashion industry including product recommendation, visual search, virtual try-on, and trend forecasting. Second-hand fashion, on the other hand, remains almost exclusively manual. The sorting and grading of second-hand clothing is a labor-intensive process that requires a lot of time and effort. The lack of automation in this sector is a major bottleneck in the transition to a circular economy.

I am fortunate to be involved in two large projects aiming to automate the second-hand fashion industry. The first project is AI for Resource-Efficient Circular Fashion funded by the Swedish Innovation Agency, Vinnova. The second project is funded by the EU: Increasing Circularity and Sustainability in Textiles and Clothing in Europe.

Sorting

Sorting is a multi-step process that involves the following steps:

• Pre-sorting: Separate shoes, household textiles like bedsheets and curtains, and other non-fashion items from the fashion items.
• Sorting fashion clothes [THIS PROJECT]: Predict various attributes of the clothing items and sort them for:
  • Reuse: Items that are in good condition and can be sold as-is. Reuse is the most sustainable option and has further sub-categories:
    • Sell in Sweden.
    • Sell outside Sweden or export.
  • Repair: Items in need of repairs, but are otherwise reusable.
  • Recycle: Items made of recyclable materials like 100% cotton.
  • Landfill: Items that are in extremely poor condition and cannot be reused or recycled.
• Fine sorting: This is the sorting that is most relevant to second-hand retailers that sell in-store and online. Their goal is to take the chunk of reusable clothing items and then decide which items to sell at what price and in what location. We do not address this in our project directly although our sorting model can be used to support this process.

In addition to this, clothes must be handled manually when they first arrive at the facility in large containers. Currently, no known technology exists that can fully automate this step although exciting advances in the field of robotics are being made.

AI-Powered Sorting

The first major challenge that any project aiming to automate the second-hand fashion industry faces is the lack of data. Existing “foundation AI models” are largely biased towards first-hand fashion since that is the kind of data that is readily available on the internet. For instance, these models are incapable of recognizing the wear and tear of second-hand clothing since first-hand fashion images are usually of pristine quality.

In the Vinnova project that I am leading from RISE, we are developing a novel dataset with 30,000 used clothing items in partnership with Wargön Innovation. The first version of the dataset has already been released:

• Dataset v1, 3000 clothing items: Zenodo link.

Important

Update February 2025: v3 with 31,638 items was released on September 19th, 2024 here. For up to date information, please visit fnauman.com/second-hand-fashion.

The dataset has been released under a permissive CC-BY 4.0 license that allows commercial use given that the authors are properly cited.

Furthermore, we are developing AI models to recognize damage on clothes and to grade them according to their quality. The scope of ongoing projects is not full automation, but to instead provide a “decision support tool”. A decision support tool is supposed to assist the human operator in making the final decision by judging the cloth condition, assessing the brand quality and how it compares with other brands in the market, and finally, estimating the best use case for the item.

Challenges and Opportunities

We have identified the following challenges and opportunities in the second-hand fashion industry:

• Data: While our dataset of 30,000 clothing items is the largest of its kind, it is still not large enough to train a deep learning model of the “foundation model” kind. Instead, we must resort to using existing foundation models and finetune them with this data. What makes this particularly challenging is that for first-hand fashion, training a model on, for example, pink T-shirts and black pants is sufficient, but for second-hand fashion, one must be able to distinguish between a pink T-shirt that is in good condition and one that is in poor condition. In other words, we need a dataset large enough to contain different degrees of damage to clothes. One major problem with lack of data will be addressed by the introduction of the digital product passport that aims to preserve the data about a product throughout its life cycle.
• Annotations: Similar to the subjectivity of language annotations, the annotations of second-hand clothing items are often specific to the annotators and the scope of the facility they are working for. For instance, Wargön Innovation works with the Swedish Red Cross and does not directly price the clothing items. In contrast, other sorting facilities like Myrorna and Björkåfrihet price the items to be sold in their own stores. This means that the annotations are not only subjective, but also specific to the business model of the sorting facility.
• Robotics: The second-hand fashion industry is still almost exclusively manual. With the recent advances in robotics, there is an exciting opportunity to fully automate the entire sorting process from the pre-sorting step to the fine sorting step.

Conclusion

The second-hand fashion industry is ripe for disruption. With the increase in global awareness about the environmental impact of the textile industry, the second-hand fashion retail is expected to grow exponentially. Nonetheless, the industry is still largely manual and lacks large scale datasets and AI models. The introduction of the digital product passport and extended producer responsibility are likely to accelerate the automation of the second-hand fashion industry. Most players in this sector are volunteer run small businesses that lack the resources to invest in AI and robotics. With project like ours, we hope to make the technology accessible to all players.