Endpoint deployment of a computer vision example

[pietz]

I would like to see an example of how a computer vision based model (like object classification or detection) can be deployed to a real-time online endpoint. Specifically, I want to know the best practices and options I have on how to handle the binary input. The following questions revolve in my head:

• Should I provide images by URL?
• Should I provide images by base64?
• Should I provide images as uncompressed pixel tensors?
• Can I provide images as multipart / form-data? (my preferred option)
• Where and how should I handle preprocessing?
• How does it nicely integrate with MLflow

Thanks