Unify image and video annotation feature vectors #1348
Description
For a LabelBOT search it should not matter if an annotation was made on an image or video. Also with Largo similarity sorting, filtering can just be done by the volume ID(s) and not the annotation type. But right now we have to choose between image and video annotations for LabelBOT (or do two queries), which may miss a lot of useful data.
But it's difficult to merge the image_annotation_label_feature_vectors and video_annotation_label_feature_vectors tables because they rely on foreign key constraints (id and annotation_id) to automatically delete vectors when an annotation (label) is deleted. Is there a good solution for that?
Regarding the vector index it should be ok to just insert all the video annotations into the image annotations table and then rename the image annotations table. We don't have to rebuild the entire index and we have much more image annotations than video annotations. If it's different in other instances we could even implement a check and insert the smaller table in the larger table.