An Active Semi-Supervised Learning for Object Detection

Abstract

The rapid progress in deep learning technology has rendered extensive annotated datasets indispensable for augmenting algorithmic performance. However, annotating datasets requires a significant amount of resources and manpower. To address the high costs of annotating object detection tasks, we present an active semi-supervised learning algorithm framework tailored for object detection. The framework utilizes a semi-supervised learning model equipped with active learning strategies to manually label difficult-to-process minority samples. Additionally, we introduce the SimOTA(Simplified Optimal Transport Assignment) label assignment strategy, which obtains optimal samples from a global perspective. Moreover, the loss function for the unlabeled data has been adjusted to enable the utilization of semantic information from the noisy pseudo labels. The empirical findings obtained from assessments conducted on publicly accessible datasets provide evidence that the active semi-supervised learning algorithm framework proposed in this paper outperforms current advanced active learning strategies and semi-supervised learning algorithms in the area of object detection.