About this document
Dense Contrastive Learning for Vision Tasks by songmhor is a document available to read on EtoBox.
The document presents Dense Contrastive Learning (DenseCL), a self-supervised learning method designed for dense prediction tasks in computer vision, addressing the limitations of existing methods that focus on image classification. DenseCL operates at the pixel level, optimizing a pairwise contrastive loss to improve performance in tasks such as object detection and semantic segmentation, significantly outperforming the MoCo-v2 baseline. The method demonstrates superior results with minimal computational o
- Author
- songmhor
- Language
- EN