Case Study
Visual Search on
Willhaben.at
Company |
Industry |
 |
E-commerce |
Client
Willhaben is Austria's largest free online marketplace, with more than 10 million currently listed items. Each day, approximately 400,000 ads are posted on the platform. The marketplace covers a diverse array of categories, including real estate, automobiles and motorcycles, job listings, and a miscellaneous section aptly named "Marketplace."
Notably, the fashion category, which includes clothing, shoes, and accessories, stands as one of the largest segments within the miscellaneous categories. The Willhaben website ranks among Austria's most visited websites, attracting over 8 million unique users on a monthly basis.
Impact
With our Visual Search solution, we helped Willhaben, Austria's largest online marketplace, boost user experience & engagement. Our solution allowed their users to find items effortlessly by using their phone camera as a search tool, enhancing the overall product search & discovery process.
Problem
Visual Search is the system that uses an image taken by the user and uses that image to search for similar items from Willhaben's active listings. Users can provide an image by taking a photo with their mobile phone, taking a screenshot on their phone, or taking a photo from a catalog.
Once the photo is uploaded and processed, the system will show a list of the most visually similar items (sorted by similarity) that can be additionally filtered by attributes such as location, price, etc.
Challenges
Scalability
The primary challenge of this project was to develop an ML algorithm capable of handling an immense volume of image-based searches. The final product needed to support millions of daily searches on Willhaben, one of Austria's most visited websites
Human Similarity Judgement
The ML-based system was required to provide item recommendations based on human similarity (factors that people find similar) rather than just color, pattern, and pixel values
Solution
First, a large classification neural network on Willhaben data is trained. Then, this network is used as a backbone for creating a system that converts an image to a numerical vector. These vectors are created so that semantically similar images are converted into similar vectors. Vectors are compared using hamming distance to enable fast comparison.
ElasticSearch is used as a search engine. All ads are put there with their vectors. Regular Elasticsearch filters can be then used to filter for price, location, category, etc., while vector similarity is used to sort the resulting items by visual similarity.
Tools and Technologies
Python
PyTorch
Elasticsearch
Falcon
Numpy
Triton Inference Server
Results
This system allows users to search the marketplace in a new and efficient way, making it easier to find the items they want. It is particularly helpful for fashion, antiques, and furniture categories where describing the style and appearance can be challenging.
Discover More Case Studies
Discover the impact of our custom AI solutions on business success through customer stories
