Semi-Supervised Object Detection with Uncurated Unlabeled Data for Remote Sensing Images

Abstract

Annotating remote sensing images (RSIs) poses a significant challenge, primarily due to its labor-intensive nature. Semi-supervised object detection (SSOD) methods address this challenge by generating pseudo-labels for unlabeled data, assuming that all classes present in the unlabeled dataset are also represented in the labeled data. However, realworld scenarios may lead to a mixture of out-of-distribution (OOD) samples and in-distribution (ID) samples within the unlabeled dataset. In this paper, we extensively explore techniques for conducting SSOD directly on uncurated unlabeled data, termed Open-Set Semi-Supervised Object Detection (OSSOD). Our approach begins by utilizing labeled in-distribution data to dynamically construct a class-wise feature bank (CFB) that captures features specific to each class. Subsequently, we compare the features of predicted object bounding boxes with the corresponding entries in the CFB to calculate OOD scores. We design an adaptive threshold based on the statistical properties of the CFB to accommodate different classes, allowing us to effectively filter out OOD samples. The effectiveness of our proposed method is substantiated through extensive experiments on two widely used remote sensing object detection datasets: DIOR and DOTA. These experiments demonstrate the superior performance and efficacy of our OSSOD approach on RSIs.

Keywords: Remote sensing images, semi-supervised learning, open-set semi-supervised learning, Object detection, memory bank, adaptive threshold