Self-Supervised Learning for Building Image Datasets

Building datasets is one of the most tedious tasks in machine learning. Image datasets consisting of unique images that are not included in large datasets require vast annotated images and bounding boxes that are created manually. We address this issue using a self-supervised learning (SSL) model for clustering before labeling the images. SSL has generally been used for pre-training. We propose the use of SSL to build image datasets by clustering using k-means. The dataset that we mainly used was CIFAR-10, and the SSL model that we used was Image Bidirectional Encoder Representations from Transformers (BERT) Pre-Training with Online Tokenizer (iBOT).