[BUG]: Image format (RGB/BGR) is inconsistent when using DeepFace.represent

[dakotah-jones]

Before You Report a Bug, Please Confirm You Have Done The Following...

• I have updated to the latest version of the packages.
• I have searched for both existing issues and closed issues and found none that matched my issue.

DeepFace's version

v0.0.93

Python version

3.12.8

Operating System

Ubuntu 20.04.6 LTS

Dependencies

N/A

Reproducible example

representations = DeepFace.represent(
    file_path,
    model_name='Facenet',
    detector_backend='skip',
    normalization='Facenet'
)

# vs

representations = DeepFace.represent(
    file_path,
    model_name='Facenet',
    detector_backend='retinaface',
    normalization='Facenet'
)

Relevant Log Output

N/A

Expected Result

I expect the image format to be consistent regardless of detector_backend.

What happened instead?

When detector_backend is "skip" the final image used for the embedding is in "RGB" format.
When detector_backend is any other valid option, like "retinaface" the final image used for the embedding is in "BGR" format.

Additional Info

When using detector_backend="skip":

• The image is loaded using image_utils.load_image. (here)
• The image is in "BGR" format.
• The image format is swapped using img = img[:, :, ::-1]. (here)
• The image is in "RGB" format.
• The image embedding is calculated.

When using any other detector_backend:

• detection.extract_faces is called.
• The color_face argument is defaulted to "rgb"
• The image is loaded using image_utils.load_image. (here)
• The image is now in "BGR" format.
• The color_face argument is "rgb" so the format is swapped using current_img = current_img[:, :, ::-1]. (here)
• The image is now in "RGB" format.
• The format is swapped again within represent using img = img[:, :, ::-1]. (here)
• The image is now in "BGR" format.
• The image embedding is calculated.