A robust training method for object detectors in remote sensing image

Abstract

With the development of convolutional neural networks (CNNs), remote sensing object detection has been made a great improvement. The CNNs-based detectors rely on accurate manually labeled training data. Due to the characteristics of remote sensing images and the professional requirements of annotation, the quality of data annotation is difficult to guarantee, and the labels of data will inevitably introduce noise. Several approaches have been developed to learn robust models from noisy labels, but most of them have focused on classification tasks. Training object detectors with noisy remote sensing data has been less investigated. In this article, we propose an object Co-teaching training strategy to train robust object detectors from noisy labels. Specifically, we train two detectors in a parallel manner, and let them teach each other to filter noise object instances in every given mini-batch. We also introduce a weighting factor to the object Co-teaching (named Reweighting object Co-teaching) which improves the performance of the detector on clean datasets without adding any complexity of the detector. Extensive experiments on DIOR and RSOD datasets demonstrate the effectiveness of the proposed method. Under 0.5 level noisy label, the object Co-teaching improves 7.63% and 9.57% mAP compared to the Baseline on the two datasets, respectively. Under clean label, the proposed Reweighting object Co-teaching improves 0.78% and 2.60% mAP on the two datasets without adding any complexity to the detector.