def facenet_template(embedding, annotation_type, fixed_positions=None):
"""
Facenet baseline template.
This one will crop the face at :math:`160 \\times 160`
Parameters
----------
embedding: obj
Transformer that takes a cropped face and extract the embeddings
annotation_type: str
Type of the annotations (e.g. `eyes-center')
fixed_positions: dict
Set it if in your face images are registered to a fixed position in the image
"""
# DEFINE CROPPING
cropped_image_size = (160, 160)
if annotation_type == "eyes-center" or annotation_type == "bounding-box":
# Hard coding eye positions for backward consistency
# cropped_positions = {
cropped_positions = dnn_default_cropping(cropped_image_size,
annotation_type="eyes-center")
if annotation_type == "bounding-box":
# This will allow us to use `BoundingBoxAnnotatorCrop`
cropped_positions.update({
"topleft": (0, 0),
"bottomright": cropped_image_size
})
else:
cropped_positions = dnn_default_cropping(cropped_image_size,
annotation_type)
annotator = MTCNN(min_size=40, factor=0.709, thresholds=(0.1, 0.2, 0.2))
# ASSEMBLE TRANSFORMER
transformer = embedding_transformer(
cropped_image_size=cropped_image_size,
embedding=embedding,
cropped_positions=cropped_positions,
fixed_positions=fixed_positions,
color_channel="rgb",
annotator=annotator,
)
algorithm = Distance()
return PipelineSimple(transformer, algorithm)
def arcface_template(embedding, annotation_type, fixed_positions=None):
# DEFINE CROPPING
cropped_image_size = (112, 112)
if annotation_type == "eyes-center" or annotation_type == "bounding-box":
# Hard coding eye positions for backward consistency
# cropped_positions = {
cropped_positions = cropped_positions_arcface()
if annotation_type == "bounding-box":
# This will allow us to use `BoundingBoxAnnotatorCrop`
cropped_positions.update(
{"topleft": (0, 0), "bottomright": cropped_image_size}
)
elif isinstance(annotation_type, list):
cropped_positions = cropped_positions_arcface(annotation_type)
else:
cropped_positions = dnn_default_cropping(
cropped_image_size, annotation_type
)
annotator = MTCNN(min_size=40, factor=0.709, thresholds=(0.1, 0.2, 0.2))
transformer = embedding_transformer(
cropped_image_size=cropped_image_size,
embedding=embedding,
cropped_positions=cropped_positions,
fixed_positions=fixed_positions,
color_channel="rgb",
annotator=annotator,
)
algorithm = Distance()
return PipelineSimple(transformer, algorithm)
def oxford_vgg2_resnets(
model_name, annotation_type, fixed_positions=None, memory_demanding=False
):
"""
Get the pipeline for the resnet based models from Oxford.
All these models were training the the VGG2 dataset.
Models taken from: https://www.robots.ox.ac.uk/~albanie
Parameters
----------
model_name: str
One of the 4 models available (`resnet50_scratch_dag`, `resnet50_ft_dag`, `senet50_ft_dag`, `senet50_scratch_dag`).
annotation_type: str
Type of the annotations (e.g. `eyes-center')
fixed_positions: dict
Set it if in your face images are registered to a fixed position in the image
"""
# DEFINE CROPPING
cropped_image_size = (224, 224)
if annotation_type == "eyes-center":
# Coordinates taken from : https://www.merlin.uzh.ch/contributionDocument/download/14240
cropped_positions = {"leye": (100, 159), "reye": (100, 65)}
else:
cropped_positions = dnn_default_cropping(
cropped_image_size, annotation_type
)
transformer = embedding_transformer(
cropped_image_size=cropped_image_size,
embedding=OxfordVGG2Resnets(
model_name=model_name, memory_demanding=memory_demanding
),
cropped_positions=cropped_positions,
fixed_positions=fixed_positions,
color_channel="rgb",
annotator="mtcnn",
)
algorithm = Distance()
from bob.bio.base.pipelines import PipelineSimple
return PipelineSimple(transformer, algorithm)
def vgg16_oxford_baseline(annotation_type, fixed_positions=None):
"""
Get the VGG16 pipeline which will crop the face :math:`224 \\times 224`
use the :py:class:`VGG16_Oxford`
Parameters
----------
annotation_type: str
Type of the annotations (e.g. `eyes-center')
fixed_positions: dict
Set it if in your face images are registered to a fixed position in the image
"""
# DEFINE CROPPING
cropped_image_size = (224, 224)
if annotation_type == "eyes-center" or annotation_type == "bounding-box":
# Hard coding eye positions for backward consistency
# cropped_positions = {
cropped_positions = {"reye": (112, 82), "leye": (112, 142)}
if annotation_type == "bounding-box":
# This will allow us to use `BoundingBoxAnnotatorCrop`
cropped_positions.update(
{"topleft": (0, 0), "bottomright": cropped_image_size}
)
else:
cropped_positions = dnn_default_cropping(
cropped_image_size, annotation_type
)
annotator = MTCNN(min_size=40, factor=0.709, thresholds=(0.1, 0.2, 0.2))
transformer = embedding_transformer(
cropped_image_size=cropped_image_size,
embedding=VGG16_Oxford(),
cropped_positions=cropped_positions,
fixed_positions=fixed_positions,
color_channel="rgb",
annotator=annotator,
)
algorithm = Distance()
return PipelineSimple(transformer, algorithm)
def afffe_baseline(
annotation_type,
fixed_positions=None,
memory_demanding=False,
device=torch.device("cpu"),
):
"""
Get the AFFFE pipeline which will crop the face :math:`224 \\times 224`
use the :py:class:`AFFFE_2021`
Parameters
----------
annotation_type: str
Type of the annotations (e.g. `eyes-center')
fixed_positions: dict
Set it if in your face images are registered to a fixed position in the image
"""
# DEFINE CROPPING
cropped_image_size = (224, 224)
if annotation_type == "eyes-center":
# Hard coding eye positions for backward consistency
cropped_positions = {"leye": (110, 144), "reye": (110, 96)}
else:
cropped_positions = dnn_default_cropping(
cropped_image_size, annotation_type
)
transformer = embedding_transformer(
cropped_image_size=cropped_image_size,
embedding=AFFFE_2021(memory_demanding=memory_demanding, device=device),
cropped_positions=cropped_positions,
fixed_positions=fixed_positions,
color_channel="rgb",
annotator="mtcnn",
)
algorithm = Distance()
from bob.bio.base.pipelines import PipelineSimple
return PipelineSimple(transformer, algorithm)