Instance-Semantic Self-Supervised Learning for Object Detection in Remote Sensing

The paradigm of conducting downstream fine - tuning after large-scale pre-training has been widely applied in various object detection algorithms in remote sensing. However, the datasets used in the pre - training and fine-tuning stages of existing large models are usually different. Especially in object detection tasks, most methods involve pre-training on classification datasets and then fine - tuning with object detection datasets. This difference in data features often causes the pre-trained weights to deviate to some extent from the actual downstream tasks. To ensure the consistency of data features in remote sensing object detection tasks, this paper proposes an Instance-Semantic Self-Supervised Learning (ISSL) framework for object detection in remote sensing. It aims to enable the instance branch and the object semantic branch to learn the spatial and semantic features of objects respectively through contrastive learning, and complete the pre-training on remote sensing object detection datasets. The framework is proposed specifically for remote sensing object detection tasks, eliminating the differences arising from the inconsistency between the pre- training dataset and the dataset for downstream task. This paper evaluates the ISSL on the DOTA and DIOR dataset and compares it with several commonly used pre-training methods. The results show that this method performs better in object detection tasks, and the performance improvement is more significant under the condition of less labeled data.