JiaRenChang / PSMNet

Pyramid Stereo Matching Network (CVPR2018)
MIT License
1.42k stars 423 forks source link

question about size #13

Open zyfsa opened 6 years ago

zyfsa commented 6 years ago

hello,is this a cvpr2018 poster paper or arxiv paper? Besides, when you test, your input is not crop(256512), is 3841248. If it will be out of memory? if my code will be out of memory when 384*1248, how can I test? thank you very much.

JiaRenChang commented 6 years ago

@zyfsa Hi, this paper has been accepted to CVPR2018. It spend around 4500 MB GPU memory to test KITTI. You may need a GPU with higher capacity if the code is out of memory

thias15 commented 6 years ago

In the latest version of the code the crop size is 1232x368 for KITTI. Shouldn't it be 1248x384?

JiaRenChang commented 6 years ago

@thias15 We pad the width and height of images to the multiple of 32.

leonzfa commented 5 years ago

训练集和测试集图像尺寸不一样对测试结果影响大吗?例如SceneFlow数据集上图像都填充到544*960是不是性能更好?

Chenlei6 commented 5 years ago

sceneflow数据集的尺寸不是540 960吗,代码中是进行了裁剪,不是填充,怎么得到544 960的结果