Data augmentation by morphological mixup for solving Raven’s progressive matrices

Raven’s progressive matrix (RPM) is one kind of visual abstract reasoning tasks, which tests the ability extracting rules from limited samples and applying them to an unknown setting. It frequently used in evaluating human intelligence. Recent advances RPM-like datasets solution models partially address challenges visually understanding RPM questions logically missing answers. This paper tackles poor generalization performance due insufficient datasets. To problem data for precisely conducting relational RPMs, we propose effective scheme, namely candidate answer morphological mixup (CAM-Mix). CAM-Mix serves as a augmentation strategy by gray-scale image mixup, regularizes various methods overcomes model overfitting problem. Compared with existing methods, more accurate decision boundary could be defined creating new negative answers semantically similar correct Experimental results show that proposed method on state-of-the-art can provide significant consistent improvements compared other strategies.