CMID A Unified Self-Supervised Learning Framework for Remote Sensing Image Understanding

• 论文名称: CMID: A Unified Self-Supervised Learning Framework for Remote Sensing Image Understanding
• Link: https://arxiv.org/abs/2304.09670
• Published in: TGRS 2023
• Type: Pretrain
• Code/Project: https://github.com/NJU-LHRS/official-CMID
• 备注: 自监督，通过蒸馏将对比学习和掩码结合在一起
• Backbone: Transformer, CNN
• Backbone 1: architecture-agnostic(CNN/Transformer)
• 下游任务: Classification, Semantic Segmentation, Object Detection, Change Detection
• 下游任务 1: scene classification, semantic segmentation, object detection, and change detection
• Short Summary: 对图片一个做增强一个做mask，经过网络后同时计算两个分支的contrastive loss和mask分支的regression loss。