geomat0101 / mediamgr Goto Github PK

If two jpgs are the same image but differ in metadata or something else not impactful then we get the same face_identifier hash computed for a face in each image, causing a collision.

Need to move away from relational FK media_id on the Faces document and replace it with an edge relationship so we can graph the same face contained by multiple media docs

assuming not a duplicate or banished content:

• get exif metadata from within the image
• technical properties about the file (hash, location, etc)
• figure out how to safely resize so facial recognition gets best quality without blowing out the memory on the GPU
• get faces

Discussed in #10

need something that mediamgr.models.connect can check to see if the schema needs to be updated.

• set version in mediamgr.schema at either module or preferably collection and graph levels
• when connect() creates the schema, it stashes the version somewhere in a special arangodb state collection
• future connects compare schema versions and if different the schema is re-applied via collection.configure and indexes get checked

at the point where we need regression suites for the model stuff

There was a noticeable difference in face recognition accuracy in some images that I ran down to the fact they were stored natively in landscape mode with an exif 'orientation' value that hints which way is up to image viewers. The recognizer was absolutely terrible on these rotated images until I re-saved them in portrait orientation using gimp, then they worked great.

dlib can probably do the rotation in memory prior to passing it off to the face detector we just need to check the exif data and see if it's something that needs doing