According to the definitions in paper (Section 5.2) , we divide the generalization errors in testing set into 4 categories:
Type I error agreeable confusion pair: Two inter-class samples confuse with each other and human cannot distinguish them
Type II error disagreeable confusion pair: Two inter-class samples confuse with each other but human can distinguish them
Type III error agreeable mis-similar pair: Two intra-class samples far from with each other and human agree they look very different
Type IV error disagreeable mis-similar pair: Two intra-class samples far from with each other and human think they are visually similar
Based on these definitions, we show some examples from each dataset (model is trained with Proxy-NCA++).
Left and Middle: A confusion pair belonging to different class. Right: the nearest neighbor for left image within the same class
Left and Middle: A confusion pair belonging to different class. Right: the nearest neighbor for left image within the same class
A mis-similar pair belonging to same class
A mis-similar pair belonging to same class
Left and Middle: A confusion pair belonging to different class. Right: the nearest neighbor for left image within the same class
Left and Middle: A confusion pair belonging to different class. Right: the nearest neighbor for left image within the same class
A mis-similar pair belonging to same class
A mis-similar pair belonging to same class
Left and Middle: A confusion pair belonging to different class. Right: the nearest neighbor for left image within the same class
Left and Middle: A confusion pair belonging to different class. Right: the nearest neighbor for left image within the same class
A mis-similar pair belonging to same class
A mis-similar pair belonging to same class