CODEWITHSUNDEEP
deep-learningpytorch
Jonghyuk Kim
Jonghyuk Kim

Reputation: 15

What is the benefit of random image crop on Convolutional Network?

I`m studying about transfer learning with the pytorch tutorial. I found pytorch tutorial author uses the different approach to train set and validation set. 

data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}

As above, training session uses the Randomly cropped image on training. But I found that after that transformation, some of the images are totally cut off the target object and remains no useful feature for detection. I am curious why the author has chosen that method, even if it is more difficult to define different pipeline to train&validation dataset. Is there any benefit of the random cropping?

Upvotes: 1

Views: 1358

Answers (1)

Benedict K.
Benedict K.

Reputation: 854

Couple of ideas behind random cropping:

In short:

  • Extending the amount of data for training
  • Making NN more robust

More detail:

The semantics of the image are preserved but the activation values of the conv net are different. The conv net learns to associate a broader range of spatial activations with a certain class label and improves the robustness of the feature detectors in conv nets.

Look at this excerpt https://link.springer.com/chapter/10.1007/978-3-319-31293-4_25

Upvotes: 3

Related Questions