AI for Second-Hand Fashion

Project overview

The second-hand fashion market is rapidly growing, fuelled by environmental concerns and a shift in consumer habits away from fast fashion. In response to this trend, our external partner TRUSS aimed to create an image catalogue for the second-hand market, facilitating the search for and comparison of different garments. However, the large variety of images poses a challenge, as it is difficult to use a single image as a 'class identifier'. Therefore, it was essential to add more images to the database to enhance its diversity and increase the retrieval performance. In this project, we proposed and implemented a straightforward pipeline that identifies images to be added to the database, with the goal of maximising performance and minimising search time.

The challenge

Platforms that facilitate the buying and selling of second-hand clothing must manage extensive databases containing user-uploaded images. These images showcase a wide variety of visual differences, which can be seen both within individual garment categories (e.g., various angles, lighting conditions, and signs of wear) and across different types of garments. This diversity poses a significant challenge for standard retrieval systems, making it difficult to return truly similar items based on image queries. Consequently, the performance of creating a vector database is often low when relying on only one image per garment, even when using vectors generated by a visual transformer explicitly trained with fashion images.

Our approach

An alternative to improve performance is to increase the number of images added to the database. This increases the space of possible images, enhancing the probability of retrieving an image of the same garment (due to similarity), but at the same time, it increases the search time as the number of comparisons increases.

This project tackled the image retrieval challenge by reducing the size of the image database while maintaining retrieval accuracy. We used large, pre-trained visual models to convert garment images into feature representations, which were then stored in a vector database. Instead of relying on every uploaded image, the proposed pipeline selects representative samples of each garment using clustering techniques, which capture internal variability by identifying subgroups within each garment category. To improve the quality of these clusters, the proposal also removes outlier images before selection. This two-step process of selecting representatives and removing noise significantly lowers computational costs without sacrificing performance, making it well-suited for scalable and sustainable image retrieval contexts.

Outcomes

We made several improvements in the retrieval pipeline:

• We significantly improved image retrieval performance. Our proposal utilised only 10% of the images in the dataset, and achieved the same or better performance as when using 100% of the images in the database.
• We adjusted the algorithm so that it could accept not only a single image for searching but also a set of images (a listing), further enhancing its performance.
• Suggestions that were made for modifications to the retrieval system to enhance its efficiency were well received by TRUSS.

Testimonial

"The help from CAMaCS has allowed us to break ground in a way that wouldn’t have been possible with our internal team alone. This shows the great results that can come from mixing the flexible nature of a start-up with the deep thinking of research. We hope that the improvements to visual search made here will make second-hand a more accessible market for all via the product lookup functionality. This promotes sustainability whilst saving people money, a win-win for all!
- Connor Mattinson, CTO, TRUSS

Impact

To the industry, the work demonstrates how strategic pruning and selection can achieve a favourable performance-efficiency trade-off, maintaining near-optimal accuracy while significantly reducing computational costs. TRUSS implemented the search in their internal pipeline, improving the retrieval performance at an extra cost during the search. As discussed with TRUSS, this development can be helpful in other large-scale image retrieval or classification tasks facing similar data challenges.

From an academic point of view, we summarised our work in a paper that was submitted and accepted at PAIS-2025, the largest showcase of real applications using AI technology.

Next steps

We continue our conversations with TRUSS and are looking for new directions and funding sources to continue working together.

For more info