Inspiration
Re:Style was inspired by the increasing demand for sustainability within the fashion industry.
We believe that something as simple as the clothes we wear every day can have a significant impact on our lives and the world around us. With this in mind, we saw an opportunity to bring a novel solution to the ever growing issues of waste and sustainability. Our platform would not only empower individuals to make eco-conscious choices but also strengthen communities by encouraging the exchange and reuse of clothing.
What it does
Re:Style employs a custom-trained AI model to digitize your wardrobe. Users regularly take photos of their outfits, which are then analyzed to build a digital wardrobe. By tracking and analyzing your preferences, the platform provides personalized outfit recommendations and offers insights into which clothing items can be donated or recycled. We support local communities by providing information on nearby charity shops for donations, all accessible directly within the app.
How we built it
We used multiple different AI models to extract information on the user's wardrobe from a set of photos:
- Fine Tuned YOLOv8 on the Deepfashion2 dataset for image segmentation.
- CLIP and BLIP models for identifying and categorizing items in the digital wardrobe.
- OpenAI GPT 4 for outfit recommendations & feature extraction.
We connected these using a Python Flask server to a React web application, with a sleek user interface for a nice user experience. Behind the scenes, we also use a Firebase database to store wardrobe information, as well as interact with the Google Maps and OpenStreetMap APIs query local charity shops.
Challenges we ran into
During the hackathon, we faced challenges with image segmentation and masking. Initially, we invested significant time in training a Mask R-CNN model on the DeepFashion2 dataset. However, the absence of a GPU on our team made the process time-consuming and inefficient. We eventually resolved this issue by switching to a customized YOLOv8 model, which provided the performance and accuracy we needed.
Additionally, we encountered difficulties achieving accurate results with CLIP and BLIP models. These models often confused similar clothing items—for example, distinguishing between trousers and shirts was manageable, but differentiating between cargos and jeans proved challenging. After several iterations, we successfully ran a larger model that provided sufficient context, ultimately overcoming the problem.
Accomplishments that we're proud of
Our team is extremely proud of the volume of working code that we were able to produce in such a short time. Finally seeing all the components of our project seamlessly integrated together was a highly rewarding experience!
We are especially proud of the speed and acuraccy with which we can extract, categorize and identify the same item of clothing from a variety of different photos.
What we learned
During the hackathon, we learned about AI models and the intricacies of training them. We deepened our understanding of image extraction, web development, and for the two of us who were first-time hackers, discovered the rewarding and often stressful experience of hackathons. Above all, we learned about the significance of sustainability and how even something as simple as our clothing choices can make a meaningful impact on a community.
What's next for Re:Style
We certainly plan to continue Re:Style in the future, growing it into a impactful force for good. This includes:
- Partnering with thrift stores, allowing users to explore nearby shops. This will be a simbiotic relationship, allowing users to buy clothes from stores, and stores to post donation requests. This feature also complements AI powered recommendations, possibly allow recommendations outside of your personalised wardrobe.
- Outfit repurposing would see the AI provide suggestions on upcycling or restyling unused or underused clothing to reduce waste and foster a more sustainable approach to dressing.
We would like to extend our heartfelt gratitude to ACSS and Majestic for making this hackathon possible. We had a lovely time at the event and hope to see you all again next year! ~ EON Team
Log in or sign up for Devpost to join the conversation.