Not sure about the other newly introduced bugs as mentioned here, but I think the face recognition thing is not a newly introduced bug. This can be verified by running the same image in certain pervious version.
The three items you mentioned are actually two things.
Thing 1:
This is expected, not bug. We use grey boxes to indicate low confidence face detections. Yellow boxes to indicate high confidence face detections. By default, auto pilot, only yellow box ones get processed. Grey ones are still shown, and clickable, in order to benefit users fast select any missing faces.
The core thing to understand here are: 1) face recognition model (industry or research) can never be perfect, such that we give some extra predictions to users to avoid missing; 2) and the model output is probabilistic (box is face = 66.67%, instead of box is face = either 0 or 1), so we take advantage to provide some low probability predictions. This feature is explained one past release I think around 3 or 4 month ago.
Thing 2:
The reason here is face size is bigger than face recognition model’s perceptual field. This problem is solvable, by training a model that has larger perceptual field, so that the model can see a whole face, rather than a part of a face (and think one face to be multiple faces). However, there is another thing, face refinement model is better at recovering smaller faces, and large faces tend to get better results with base filters (eg sharpen).
To conclude, we know shipping another bigger face recognition model can get better recognition on bigger faces; but those bigger faces get better results without face refinement model. Meanwhile, shipping another face recognition model can take extra space, longer runtime. Considering the above, we did not ship the model/decided to not to ship the fix.
Of course, one single lightweight model recognition model that can detect faces of any sizes is the perfect solution, but that’s need another round of model iteration, and there are many technical challenges (each year top computer vision publications still see many new face recognition related papers trying to solve a lot of those challenges).
