MEJIGCLU: More Effective Jigsaw Clustering For Unsupervised Visual Representation Learning

Abstract

Unsupervised visual representation learning aims to learn general features from unlabelled data. Early methods design intra-image pretext tasks as learning targets and can be achieved with low computational overhead but unsatisfactory performance. Recent methods introduce contrastive learning and achieve surprising performance, but multiple views of training data are required in one batch, resulting in high computational overhead. To achieve competitive results to contrastive learning with low computational overhead, we propose a new unsupervised representation learning method with jigsaw clustering and classification as pretext tasks motivate the network to learn discriminative feature. To increase the data diversity, we propose to partition each training image into patches with random overlap, then randomly permute and stitch them into new training batch. Comparing with SOTAs, our method achieves state-of-the-art performance on both image classification/semi-classification on ImageNet and object detection on COCO.