Using AI to Find Your Celebrity Stylist (Part II)

In this tutorial, we extended our first celebrity-style project by using Milvus' new dynamic schema, filtering out certain segmentation IDs, and keeping track of the bounding boxes of our matches. We also sorted our search results to return the top three results based on the number of matches. Milvus' new dynamic schema allows us to add extra fields when we upload data using a dictionary format, changing the way we were initially batch-uploading a list of lists. It also facilitated adding crop coordinates without changing the schema. As a new preprocessing step, we filtered out certain IDs that aren't clothing-related based on the model card in Hugging Face. We filter these IDs out in the get_masks function. Fun fact, the obj_ids object in that function is actually a tensor. We also kept track of the bounding boxes. We moved the embedding step to the image cropping function and returned the embeddings with the bounding boxes and segmentation IDs. Then, we saved these embeddings into Milvus using a dynamic schema. At query time, we aggregated all the returned images by the number of bounding boxes they contained, allowing us to find the closest matching celebrity image via different articles of clothing. Now it's up to you. You can take my suggestions and make something else out of it, such as a fashion recommender system, a better style comparison system for you and your friends, or a generative fashion AI app.