Review. Dense Prediction Tasks for SSL
- 1. Review : Dense Prediction Tasks for Self-SL Sungchul Kim 2022. 01. 06
- 2. Contents ▪ Introduction ▪ DenseCL (CVPR 2021), PixPro (CVPR 2021), SCRL (CVPR 2021) ▪ DetCon (ICCV 2021) ▪ MaskCo (ICCV 2021) ▪ ReSim (ICCV 2021) ▪ RegionCL (arXiv:2111.12309) ▪ SoCo (NeurIPS 2021)
- 3. Introduction ▪ Learns from unlabeled sample data • Pretext task • Contrastive SSL vs. Non-contrastive SSL • Instance discrimination vs. Dense prediction • Downstream task • classification, detection, segmentation, … Self-Supervised Learning
- 4. Introduction ▪ Instance Discrimination (image-level comparison) • treats each image in a training set as a single class Self-Supervised Learning SupCon : https://arxiv.org/pdf/2004.11362.pdf SimSiam : https://arxiv.org/pdf/2011.10566.pdf
- 5. Introduction ▪ Dense Prediction (pixel-level representation) • compares how closely each pixel or area is related Self-Supervised Learning DenseCL : https://arxiv.org/pdf/2011.09157.pdf MaskCo : https://arxiv.org/pdf/2108.07954.pdf
- 6. DenseCL (CVPR 2021), PixPro (CVPR 2021), SCRL (CVPR 2021) DenseCL : https://arxiv.org/pdf/2011.09157.pdf PixPro : https://arxiv.org/pdf/2011.10043.pdf SCRL : https://arxiv.org/pdf/2103.06122.pdf BYOL SCRL
- 14. ReSim (ICCV 2021) Pascal VOC MS COCO https://arxiv.org/pdf/2103.12902.pdf