Personal Project, 2017. 11. ~ 2018. 6.
Project Summary
Shoepy is a shopping mall application where users can search for similar shoes by images. At the time I conducted this project, I was deeply interested in deep learning. Since I wanted to apply it to developing something useful, I started this project. Overall system architecture is like below.
The most complicated thing was designing model training pipeline. Crawler had to collect new products data from shopping malls periodically. And I had to train and serve new model reflecting it to provide newest and reliable search results. Image search was implemented like below.
A shoes image is given as an input. The CNN classifier model predicts the label of the input image and extracts the feature vector of it. Using those, calculate the ANN(approximate nearest neighbors) using an index file that has undergone vector quantization in advance. Get the products data from database with ANN results and return image search response.
Role
- Developed crawler and collected over 140K shoes data from shopping malls.
- Trained convolutional neural network with custom datasets.
- Implemented vector indexing and searching server.
- Developed frontend & backend application.
Tech Stack
- Language: Java, Python
- Deeplearning Framework: Pytorch
- Approximate Nearest Neighbor: Faiss
- DB: postgresql
Results
- Demo
-
frontend application
- image search results
