Image similarity for finding close garments

Inspiration

After being introduced to recommender systems at university, it was clear that a distance metric approach could be useful for this application in particular, and so we started to look for ways to make this approach viable.

What it does

After implementing a stable UI, we can successfully recommend 5 images of clothes and articles that are more closely related to the image that is selected by clicking the button under it.

How we built it

First of all, the images are loaded and pre-processed to ensure uniformity. Then the features are extracted by using the VGG16, and ResNet50 networks. Then the dimension of the features is reduced by applying PCA. From this, a cosine similarity matrix is created, which is used to obtain the indices of the most similar items photos that are visualized. This cosine similarity is stored as a CSV file so that it can be called through the Web Application used to present the project.

Challenges we ran into

First, we wanted to train a network to find similar items, but the lack of tags meant that we couldn't compute a loss function, so a different approach was needed.

We have never worked with frontend applications and doing so in such a limited time of amount was a challenge in itself.

Accomplishments that we're proud of

The project itself is a satisfying way to implement what we have learned in our bachelor's first years, and given it provides coherent results consistently we are proud of being able to apply our knowledge in such a direct way.

Also, we are proud of our first front-end application.

What we learned

We have improved our group working ability, since collaborating so closely for a common goal with the pressure of having to deliver a result before a deadline arrives, provides an amazing context to do so. We have also learned to use different libraries like TensorFlow. Furthermore, we have gained a first insight into front-end programming and its difficulties.