VGG16 preprocessing buggy?
When using the VGG16 network, we need to subtract the RGB mean from the channels. As the images are in bob format (NxCxHxW
), we would need to subtract the mean from [:,i,:,:]
. Instead, we subtract it from [:,:,:,i]
:
https://gitlab.idiap.ch/bob/bob.bio.face/-/blob/3567e990d0e523ceb5d3f9598054d8a27d7f7000/bob/bio/face/embeddings/opencv.py#L140
This is most certainly incorrect, especially since we use the correct dimension later on to convert RGB to BGR: https://gitlab.idiap.ch/bob/bob.bio.face/-/blob/3567e990d0e523ceb5d3f9598054d8a27d7f7000/bob/bio/face/embeddings/opencv.py#L146
Finally, in the pipeline, we define an MTCNN annotator with particular parameters: https://gitlab.idiap.ch/bob/bob.bio.face/-/blob/3567e990d0e523ceb5d3f9598054d8a27d7f7000/bob/bio/face/embeddings/opencv.py#L203
but this is ignored since the pipeline uses "mtcnn"
.