Adaptive Weight of Unreliable Relation Module for Semi-supervised Multi-label Image Recognition

Abstract

Semi-supervised multi-label classification is a challenging task due to the insufficient training guidance and unknown label co-occurrence probabilities. For papers in recent years, relation modules are widely utilized to explore the potential label relationships, but the severe frequency-biased issue between a global relationship and local images significantly degrades their effectiveness. Besides, the difference in data distribution between training and testing sets further affects the performance, especially when the labeled data is limited. To address these problems, we propose a simple selective relation module to learn an adaptive weight of relation module for each image and enforce the consistency between relation-based predictions and initial predictions. In addition, we exploit the augmentation-based consistency loss to generate more confident relation-based pseudo-labels and more robust relation-importance predictions. Our methods can be added to a variety of relation-based multi-label classification methods and we show our improvements in the classification accuracy on Pascal VOC dataset.