Exploring Unsupervised Learning Paradigms for Object Detection: A Survey

Unsupervised visual representation learning, also known as self-supervised learning, seeks to derive features without relying on manual annotations. Unsupervised learning algorithms for object detection are receiving increasing research attention as an alternative to resource-intensive supervised learning approaches requiring extensive manual annotations. This survey provides a comprehensive review of recent unsupervised and self-supervised learning paradigms proposed for object detection in images and videos. The key unsupervised learning approaches examined include contrastive learning, region proposal methods, clustering, adversarial training, and fusion of imperfect saliency models. The surveyed techniques incorporate innovative pretext tasks, multi-view comparisons, student-teacher frameworks, and noise modelling to enable unsupervised feature learning for detection. Evaluations demonstrate state-of-the-art performance on benchmark datasets surpassing supervised models in some cases. The survey provides an in-depth analysis of relative merits, limitations, and promising future directions to advance unsupervised object detection.