Multitask Regularization for Image Aesthetic Evaluation

Convolutional neural networks are data hungry and in cases when annotation is costly or difficult, additional information from other sets may be welcomed. In this paper, to improve the performance on the main task, we introduce a secondary one, over unlabeled data to provide better structuring. The solution falls in the theme of multiple task learning and unlabeled data is integrated next to the main regression task by classification via pseudo-labeling. The method is showed to improve the baseline performance for image aesthetic assessment on the AADB benchmark.