Key Points

• Basic strategy: self-supervised learning to reduce reliance on human-labeling
• Downstream tasks: dense prediction (segmentation, depth estimation, etc.)
• Dataset: Pascal VOC 2012
• Baselines: MoCo and SimCLR with semantic segmentation head (a decoder)
• Main effort: design contrastive pretext tasks for segmentation

• Possible direction: auto-encoder and energy-based method (why or why not)

  To-do’s

• [x] Intro to MMSegmentation toolbox

• Convert MoCo & SimCLR model checkpoints to MMSegmentation
• Related work about auto-encoder style self-supervised learning