1. Image Classification

1-1. Receptive field in CNN

• The region in the input space that a particular CNN feature is looking at
• Ideally, we would like each output to have a big receptive field, so as to ensure that no crucial information was not taken into account.

“We observe a logarithmic relationship between classification accuracy and receptive field size, which suggests that large receptive fields are necessary for high-level recognition tasks, but with diminishing rewards.” - Araujo et al(2019)

2. Annotation, Data Efficient Learning

2-1. Data augmentation

2-1-1. Learning representation of dataset

• Dataset is (almost) always biased

• The training dataset is sparse samples of real data

• Datasets do not fully represent real data distribution!

   →Augment data to fill more space and close the gap
  

2-1-2. Data augmentation

• Brightness adjustment
• Rotate, flip
• crop
• Affine transformation (shear)