Object-Centric Masked Image Modeling-Based Self-Supervised Pretraining for Remote Sensing Object Detection

Masked image modeling (MIM) has been proved to be an optimal pretext task for self-supervised pretraining (SSP), which can facilitate the model capture effective task-agnostic representation at step and then advance fine-tuning performance of various downstream tasks. However, under high randomly masked ratio MIM, scene-level MIM-based SSP is hard small-scale objects or local details from complex remote sensing scenes. Then, when pretrained models capturing more information are directly applied object-level step, there obvious learning misalignment between steps. Therefore, in this article, a novel object-centric (OCMIM) strategy proposed make better further object detection step. First, learn involving full scales multicategories SSP, data generator automatically setup targeted according themselves, provide specific condition pretraining. Second, attention-guided mask designed generate guided MIM task, lead discriminative highly attended regions than by using masking strategy. Finally, several experiments conducted on six benchmarks, results that OCMIM-based way normally used methods.