use swin_b in different models comparisons.
-
training set (N=20000)
-
val set (N=1250)
-
test set (N=1250)
After observation, there are some major types of fail cases:
- Not dogs and cats
- Ambiguous label, there are dogs and cats shown in the same image
- Cartoon style
- Low resolution or object was too small