摘要: Recent advances in self-supervised contrastive learning yield good image-level representation, which favors classification tasks but usually neglects pixel-level detailed information, leading to unsatisfactory transfer performance... 展开 Recent advances in self-supervised contrastive learning yield good image-level representation, which favors classification tasks but usually neglects pixel-level detailed information, leading to unsatisfactory transfer performance to dense prediction tasks such as semantic segmentation. In this work, we propose a pixel-wise contrastive learning method called CP~2 (Copy-Paste Contrastive Pretraining), which facilitates both image- and pixel-level representation learning and therefore is more suitable for downstream dense prediction tasks. In detail, we copy-paste a random crop from an image (the foreground) onto different background images and pretrain a semantic segmentation model with the objective of 1) distinguishing the foreground pixels from the background pixels, and 2) identifying the composed images that share the same foreground. Experiments show the strong performance of CP~2 in downstream semantic segmentation: By finetuning CP~2 pretrained models on PASCAL VOC 2012, we obtain 78.6% mIoU with a ResNet-50 and 79.5% with a ViT-S. 收起