search

Haochen-Wang409 / U2PL

[CVPR'22 & IJCV'24] Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels & Using Unreliable Pseudo-Labels for Label-Efficient Semantic Segmentation
Apache License 2.0
437 stars 61 forks source link

对compute_unsupervised_loss计算thresh的疑惑 #151

Closed wade0604 closed 1 year ago

wade0604 commented 1 year ago

您好！ 请问在compute_unsupervised_loss这个函数中为什么thresh是对一个batch内所有图片一起算的，而不是对单张图片分开计算thresh？

Haochen-Wang409 commented 1 year ago

一个 batch 内算的 thresh 会更加准确，对于 reliable/unreliable 的划分更加精细。