Powering a Visual Search System with Image Embedding

Highlights from the current model:

• Reduced index size by 85% over baseline model (163.8 MB to 25 MB) through dimensionality reduction using an autoencoder network
• Increased mAP score by 9% over baseline model through strategic adjustments, including increasing convolutional and max-pooling layers in the network architecture.

I have also created a blog post underlining the process, evaluation, analysis, visualization, recommendations, and sample results on this project, available here: https://www.joankusuma.com/post/powering-visual-search-with-image-embedding

The dataset used to train the model is available here:

    @online{Eileen2020,
  author       = {Eileen Li, Eric Kim, Andrew Zhai, Josh Beal, Kunlong Gu},
  title        = {Bootstrapping Complete The Look at Pinterest},
  year         = {2020}
}