Meeting Summary

1. Confirmed that pre-training decoder is beneficial.
2. Continue to explore decoder pre-training with more complex decoders.
3. Re-emphasis goals
   1. 预训练模型在segmentation的表现出众
   2. 预训练方式能够用很少的computation和data达到comparable的效果
   3. 预训练在label极少的情况下好于现有方法，极限情况为few-shot segmentation
   4. 预训练在其他dense prediction tasks（keypoint detection and depth prediction）上表现良好

      To-do’s

• 更复杂的decoder
• 整理大表，下周meeting总结