Weekly Summary

最近一次meeting总结：https://www.yuque.com/docs/share/803a018c-84e4-42c1-8e1a-ec4b6fdf4412?# 《20210105》

1. PSPHead did not perform well as expected. Limited time to do full ablation study on this.
2. 对于MoCo v1和v2，PLEDMoCo pre-training的表现相似。
3. Thoughts on using auto-encoder loss to retain detail information.

   Test Report

   MoCov1 v.s. MoCov1 + PLEDMoCo(VOCaug)

• Performance comparison
  • MoCo v1 + PSPHead: 75.45%
  • MoCo v1 & PLEDMoCo(voc-20ep, lr-0:1): 73.41%
  • MoCo v1 & PLEDMoCo(voc-20ep, lr-0:1): 73.76%
• Learning rate schedule:

MoCov2 v.s. MoCov2 + PLEDMoCo(VOCaug)

• Performance comparison
  • MoCo v2 + PSPHead: 75.06%
  • MoCo v2 & PLEDMoCo(voc-20ep, lr-0:1): 75.57%
  • MoCo v2 & PLEDMoCo(voc-20ep, lr-0:1): 75.64%
• Learning rate schedule:

一个特别的点是，在validation set上，pspnet最后的pixel mean accuracy都下降了，但loss都差不多，不知道这是否为过拟合现象。