Object detection service (#7)

[traffic1000]

Issue taken from previous Taks list (# 7)

Automatically process documents, where Object detection wasn't previously done. Service, accordingly to OCR service. Code: https://gitlab.ethz.ch/swisstopo/revaluatearchive/-/tree/master/src/object_detection\r\n\r\n

1. Save Object Meta information from Objects in table asset_object_info; all attributes should already be there
2. Save Object (or whole Object page) as separate document (image)
3. Update StatusWork of the corresponding Asset: status_work['status_work_item'] == 'objectsExtracted' (from value list) & status_work['status_work_date'] == timestamp

Unsolved/ideas:

• Possibility that we do not need to save all those jpg files. Can we reproject the ObjectBBox to the pdf file? Now we would have all Documents twice as pdf and as jpg. Create jpg on the fly?
• Manually labeled (or QC) object vs. predicted objects. Probably we need some sort of a edit possibility, that a user can change wrong predicted labels (-> generation of new training set which would then give us the option to retrain our model); some ideas on that were already made in the prototype, e.g.: https://gitlab.ethz.ch/swisstopo/revaluatearchive/-/blob/master/src/web_app/app/documents/templates/documents/page_detail.html

Update for unsolved parts:

• Save of Objects separately as image files with the file name convention: (NR=number as it its possible that two identical classified objects can occur on the same page)
• Edit possibilites: on hold, to be solved in a late stage. First we need to see if object detection works well