Offline data augmentation for rescale along the Y axis only

[utilisateur3498]

Hello,

I use TAO to train a detectnet_v2:resnet10 model. TAO is an excellent tool !

I’d like to use offline data augmentation for a special use case and I’m not sure I can do it with TAO. Perhaps I need a tool of my own.

Here is the use case:
Images of different size enter the network. They all have the same width but different height. These image contains people and the people have a “human aspect ratio” on them. Now, when the image enter the network, the people are re-scaled vertically right ? They get smaller or taller depending of the original image height.

Here is my understanding:
So I understand that I have to create a data augmentation to make the people randomly taller or smaller. But apparently I cannot zoom in or zoom out in only one direction.

Questions:
Am I right when I understand that the network will create various aspect ratio for people ?
What’s the best way to train the model with such a constraint ?

Thank you,
Best regards,
Mathieu

[Morganh]

For detectnet_v2, there is an augmentation module which provides some basic pre-processing and augmentation during training. See more in DetectNet_v2 - NVIDIA Docs.
So, usually it is not needed to generate augmented data offline.

More, there is also a tool for offline data augmentation, please refer to Offline Data Augmentation - NVIDIA Docs.

[utilisateur3498]

Hello,

Yes I read the docs but I cannot find anything related to rescale in only one dimension. I’ll do it myself. Thank you.

Mathieu