Most of the data augmentation library appears to be useless for modern vision development since it doesn’t take in bounding box and segmentation information or return the parameters of a random transform.
Will this ever be addressed?
Most of the data augmentation library appears to be useless for modern vision development since it doesn’t take in bounding box and segmentation information or return the parameters of a random transform.
Will this ever be addressed?
I’m not sure which data augmentation libraries you are referring to, so would recommend to take a look at this Detection Tutorial using torchvision
and these docs for albumentations
, which are both using bounding boxes in their transformation.
Each torchvision.transforms
provides the get_params
method, which will return the parameters used for the current setup in the transformation.
In case you are still seeing missing functionalities, I’m sure the code owners would appreciate a pull request.