Main results Figure 1 depicts the transfer loss for all methods and for all source-target domain pairs. The best transfer is achieved by the SVM trained on our transformed features in 11 out of 12 cases (SCL is only slightly better in Kitchen -> Electronics) and signficantly better for 8 cases. Interestingly,
for each target domain,
there is one case of negative transfer loss for SDAsh: an SVM trained on a different domain can outperform an SVM trained on the target domain because of the quality of our features.