weiyithu / SurroundDepth

[CoRL 2022] SurroundDepth: Entangling Surrounding Views for Self-Supervised Multi-Camera Depth Estimation
MIT License
247 stars 38 forks source link

不能复现论文结果 #19

Closed DongyangHuLi closed 1 year ago

DongyangHuLi commented 1 year ago

作者您好, 我尝试复现论文中在DDAD数据集的实验结果。但是在20个epoch中,最好的结果也达不到论文中所展示的结果。 在scale-ambiguous evaluation中,最好的结果如下: all abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 | & 0.208 & 3.474 & 12.536 & 0.309 & 0.720 & 0.886 & 0.944 \ 以下是论文的结果: image

weiyithu commented 1 year ago

您好,您可以从以下几个方面检查下:

  1. 能否用我们给的预训练好的模型复现出精度,来验证下环境是否有问题。
  2. 我们是取得最好的epoch的结果,不是最后一个epoch,epoch之间精度会有抖动。
  3. 我们是用8卡训练的,每张卡batch size为1