DFL provides something called "random warp" that distorts the images while training. This makes the model much more stable for angles and expressions that are not part of the datasets. Torchvision doesn't have something like this. Only the basic transforms. We need to implement a custom one in order to create more stable results.
DFL provides something called "random warp" that distorts the images while training. This makes the model much more stable for angles and expressions that are not part of the datasets. Torchvision doesn't have something like this. Only the basic transforms. We need to implement a custom one in order to create more stable results.