I was wondering when preprocessing steps get applied to images. Are steps like “static crop” only applied to images before training or do they get applied during inference as well? If applied at inference do both the API and Offline (Docker) modes do this?
Good question that we should document better because the answer is slightly different for each deployment method.
The hosted API does almost all of the same preprocessing with the exact same code as the data preparation step does (with tiling and static crops being the exception). We recommend resizing beforehand though because it will greatly increase the speed (due to reduced bandwidth and network latency).
Reach out to the support team for info on individual Dockers and we will look it up internally for you until we have a chance to document it!