Data

Data Augmentation

Datasets

Field of View

• Paper that is trying to address the use of best field of view: Fixing the train-test resolution discrepancy
• There is an approach to find best image patch resolution for the given datasets and tasks automatically using neural architecture search

Larger image size

• Large image size yields more accurate model in most cases
  • Model can learn global context and also semantic details
  • Given the same model, larger image resolution seems always better on image classification
  • Given the same model, larger image resolution seems always better on object detection
  • Given the same model, larger image resolution seems always better on pose estimation
• For image segmentation and given enough datasets (example: n=100) using largest image resolution possible is often the challenge winning solution
  • Authors used arbitrary resolution (e.g. 96x96) when the input image size is not static but variable. But they used largest input resolution (256x256) when it's possible.

Smaller image size

• Smaller image size yields more semantic details rather than global context of an image
• For image segmentation with relatively small datasets (example: n=20) using smallest image patch resolution is sometimes challenge winner

Medical Image Segmentation

Patch-based segmetation and limited FOV

Medical Segmentation Decathlon