Improving Self-Supervised Learning for Multi-Label Classification Using Mix-Based

Abstract

Self-supervised learning is an active area of research that had great advances in recent years. It is concerned with overcoming the bottleneck of any supervised deep learning framework that relies on hand crafted labels to achieve good results. Several State-Of-The-Art self-supervised algorithms have been proposed recently, where, most of these methods were designed to learn from ImageNet. Since ImageNet is an object-centric dataset, little focus was given to complex scene understanding tasks such as multi-label classification or object detection in the design of the self-supervised algorithm. In this work, we investigate the question of whether applying mix-based augmentations in the self-supervised pretraining would lead to better transferability to the complex downstream task of multi-label classification. “SimSiam”; a non-contrastive self-supervised algorithm; is used for pretraining along with three proposed variations of mix-based augmentations. The task of multi-label classification on PascalVOC as a non-object centric dataset is selected for evaluation of the learnt features. The SimSiam with the use of UnMix augmentation technique achieved the best performance among all our experiments with a higher mean Average Precision (mAP) than the baseline by (+3.7). We conclude that lowering the mixing probability and using the new mixed image pairs as additions to the loss function is more efficient than replacing the original images completely.