Skip to content

Latest commit

 

History

History
executable file
·
38 lines (23 loc) · 1.63 KB

setup.md

File metadata and controls

executable file
·
38 lines (23 loc) · 1.63 KB

CVP-MVSNet实验配置

环境配置

需要下载作者提供的数据集,跟MVSNet的数据集不是很一样,但都不大2G+1G

整体环境推荐不要复用MVSNet,主要是PyTorch版本比较新,再依照requirements安装下没安过的库即可

如果opencv-python在conda中无法安装成功,就直接用pip安装,一样可以用的

我最开始复用了MVSNet的cuda和torch环境,会在很多位置报错RuntimeError: output with device cuda:0 and dtype Float doesn't match the desired device cuda:0 and dtype Double,定位到报错位置将变量改为xx.double()进行格式转换会消除一些,但有些始终无法修改成功,所幸从头安装了

Train

修改下train.sh里的配置,直接运行就可以,对GPU的消耗不是很大,迭代的速度也比较快

Test / Eval

安排好数据集的位置就可以直接运行sh eval.sh进行测试即可,最终深度图结果会在outputs_pretrained

一点需要注意的是,如果运行时报错则需要指定一个GPU: CUDA_VISIBLE_DEVICES=0

eval之后会在outputs中生成每个场景的confidencedepth_est

fusion

记得要使用python2进行fusion,我最后是重新安装了一个python2的conda环境成功的

安装好环境,链接一下fusibile的位置到CVP-MVSNet/fusibile/目录下

ln -s /media/public/yan1/doublez/realdoubleZ/Developer/MVS/fusibile/build/fusibile fusion/fusibile

fusion之后会在outputs里的fusibile_fused生成每个场景的文件夹,文件夹里就是那些一大堆内容,最核心的是

  • cams/
  • images/
  • consistencyCheck/final3d_model.ply