Supplementary for Dual Progressive Prototype Network for Generalized Zero-Shot Learning

Since some brand-new methods utilize post-processing, such as calibration stacking [5] or domain detector [2, 12], to alleviate the domain shift problem, we report the results of our Dual Progressive Prototype Network (DPPN) with post-processing in Table 3 of the main paper for fair comparison. In this part, we further compare our DPPN with recent methods that clearly report their results without post-processing, of which the comparison results are shown in Table 1. APN [15] only reports their results with calibration stacking. Our DPPN outperforms the best one by respectively 15. 3%, 8. 8%, and 7. 3% for H on CUB, AWA2, and aPY datasets, and obtains comparable performance on SUN dataset. This demonstrates the effectiveness of learning representations that progressively explore category discrimination and attribute-region correspondence.