Self-supervised dense visual representation learning

In this thesis, we study on self-supervised dense visual representation learning. We propose three different approaches to be used in dense tasks such as object detection and instance segmentation. In the first one, we propose consistency regularization at both image-level and pixel-level to alleviate the false negative pairing problem in the contrastive learning. In the second, we combine two positive-pair matching strategies implemented at the pixel-level, based on similarity and distance, to eliminate the weaknesses of these strategies. In the third one, we adapt the relational self-supervised framework (ReSSL) proposed in image-level to dense-level. Contrary to common practice, we pre-trained these approaches on the COCO dataset, not on ImageNet, and evaluated them on the Pascal VOC object detection and COCO object detection and instance segmentation tasks. In our experiments, we observed that consistency regularization, using two separate pixel-level matching strategies together, and stimulating ReSSL to the pixel-level lead to an increase in performance, both quantitatively and qualitatively.