Object-Centric Masked Image Modelling for Self-Supervised Pre-Training in Remote Sensing object Detection

Abstract

The proliferation of remote sensing technologies has led to an increasing demand for effective object detection in satellite and aerial imagery, with applications ranging from environmental monitoring to urban planning. Traditional methods for analyzing such imagery often rely on manual inspection, which is both time-consuming and prone to human error. While recent advancements in automated object detection have improved efficiency, these systems frequently suffer from limitations in accurately identifying and classifying objects due to their reliance on simplistic masking techniques and insufficient context understanding. In this work, we propose a novel Object-Centric Masked Image Modelling (OCMIM) algorithm designed to enhance self-supervised pre-training for remote sensing object detection. The OCMIM algorithm comprises two key components: the Object-Centric Data Generator (OCDG) and the Attention-Guided Mask Generator (AGMG). The OCDG component empowers the model to capture comprehensive object-level context information, accommodating various scales and multiple categories, thus enriching the pre-training process. Complementing this, the AGMG focuses on improving the reconstruction of object regions by intelligently masking the most attention-worthy regions instead of employing random masking, thereby enabling more accurate object detection and classification. Our proposed OCMIM algorithm leverages the strengths of existing pre-trained models such as Mask R-CNN (M-RCNN) and RetinaNet, enhancing their performance through the integration of OCDG and AGMG. For evaluation purposes, we utilized several pre-trained models, including M-RCNN and RetinaNet, and conducted experiments on diverse datasets such as NWPU, DIAR, and UCAS. Given the extensive training time required for these models, we specifically employed M-RCNN in conjunction with OCMIM for detailed experiments on the NWPU dataset.