Use Case Retail

Retrieval fashion system

The project involves developing a fashion retrieval system using deep learning. This system is crucial for the client, who is a major player in the fashion industry, as it aims to enhance the shopping experience by allowing users to find similar clothing items through image queries. Also, it helps the commercial team to find similar clothes to display on the web.

This process involves two parts, on the one hand the detection of the clothes and on the other hand the obtaining of similar images through representation with embeddings.

Challenges

Using neural networks provides advantages over basic computer visión methods.

Create and maintain a changing dataset

Handling a continuously evolving dataset with new fashion trends, seasonal collections, and user-uploaded images requires constant updates and retraining of the model.

Unsupervised process

Trust that the images correctly represent the object

Manual Labeling Effort

The need to manually tag images for training increases the workload, making dataset preparation time-consuming and resource-intensive.

Solution

Solutions in this field usually involve the use of Siamese networks to obtain image embeddings.

In this project we went one step further, a YOLO network was trained to detect each of the garments and a personalized training was created for each of the detected clothes. To create the network, a Siamese network model was used with a particularity, the resulting embedding of the images was built by different images of a product, instead of using only one.