Detected face position sometimes inaccurate

[illnesse]

Memories has a feature that displays a rectangle around recognize's face data, see pulsejet/memories#79

I noticed that the rectangle was off sometimes so i filed this pulsejet/memories#93 But it seems like the wrong position is actually coming from recognize.

It doesn't happen often but when it does,

• it always happens to the same files (after a reset/recrawl)
• it very often happens in a row, like many pictures from the same day, a whole photo session etc. (it's always the same resolution/camera/mimetype etc)
• the file is often still correlated to the right person (in the correct cluster)
• sometimes many of those "bad" detections are clustered together
• its dimensions seem more or less correct (compared to the actual face in the image), just the position is off
• there's nothing else in the picture that could be confused with a human face

I can happily supply test data/files if needed!

Some more examples:

[reox]

My first guess was that this might happen if the image has a rotation stored in EXIF data and thus the position of the bbox is measured in the "original image space" while the image is then correctly rotated on preview but the bbox is not... But then there are images where also with a rotation it does not fit.
However, I still think that exif orientation is at least one of the problems here...

[marcelklehr]

I've had similar troubles with the v2 interface before the UI was moved to photos and it was decided to not show face indicators. All of them were related to image orientation and client-side bugs. The face detection is quite solid from my experience.

[illnesse]

Now that i know it's probably an EXIF thing, i'm going to do some more debugging and take it up with @pulsejet thanks!