medusa.crop.bbox_crop#

Module with a “crop model” which crops an image by creating a bounding box based on a set of existing (2D) landmarks.

Based on the implementation in DECA (see ../recon/flame/deca/license.md).

Module Contents#

class medusa.crop.bbox_crop.BboxCropModel(lms_model_name='2d106det', output_size=(224, 224), device=DEVICE)[source]#

A model that crops an image by creating a bounding box based on a set of face landmarks; based on the implementation from DECA.

Parameters:

  • name (str) – Name of the landmark model from Insightface that should be used; options are ‘2d106det’ (106 landmarks) or ‘1k3d68’ (68 landmarks)

  • output_size (tuple[int]) – Desired size of the cropped image

  • detector (BaseDetector) – A Medusa-based detector

  • device (str) – Either ‘cuda’ (GPU) or ‘cpu’

forward(imgs)[source]#

Crops images to the desired size.

Parameters:

imgs (str, Path, tuple, list, array_like, torch.tensor) – A path to an image, or a tuple/list of them, or already loaded images as a torch.tensor or numpy array

Returns:

out_crop – Dictionary with cropping outputs; includes the keys “imgs_crop” (cropped images) and “crop_mat” (3x3 crop matrices)

Return type:

dict

class medusa.crop.bbox_crop.InsightfaceBboxCropModel(output_size=(192, 192), scale_factor=1.5, device=DEVICE)[source]#

Crop model based on a (scaled and) square bounding box as implemented by insightface.

Parameters:

device (str) – Either ‘cuda’ (GPU) or ‘cpu’

forward(imgs)[source]#

Runs the crop model on a set of images.

Parameters:

imgs (torch.tensor) – A torch.tensor of shape N (batch) x C x H x W

Returns:

out_crop – Dictionary with cropping outputs; includes the keys “imgs_crop” (cropped images) and “crop_mat” (3x3 crop matrices)

Return type:

dict