A Distinguishable Pseudo-Feature Synthesis Method for Generalized Zero-Shot Learning

Figure 8

T-SNE visualization of features in 40 seen classes (a, b), 10 unseen classes (c, d) and all 50 classes (e, f) on AWA2 by PFS (a, c, e) and DPFS (b, d, f) in GZSL tasks. Different colours denote different classes. It is obvious that DPFS can provide more separable classes.

(a)

(b)

(c)

(d)

(e)

(f)