I am familiar with using image augmentation techniques when training image classification models and in my experience, it can greatly improve model performance especially when the dataset is small or imbalanced.
However, it seems image augmentation is not used when creating semantic search databases. My intuition says that if you’re creating a semantic search database using something like FAISS where your dataset is small - 25 K images - doing image augmentation could improve accuracy when the database is used in the real world. However, I cannot find any examples of this. Why is image augmentation not used when creating semantic search databases?
Hugging face example blog: Image Similarity with Hugging Face Datasets and Transformers
For example, if I was creating a semantic search database with images of birds than doing horizontal flipping for every dataset image