JiaRenChang / PSMNet

Pyramid Stereo Matching Network (CVPR2018)
MIT License
1.43k stars 424 forks source link

about result accuracy on the SceneFlow and the KITTI 2015 dataset #87

Open IceTTTb opened 6 years ago

IceTTTb commented 6 years ago

您好,我想请问一下您在paper中列出的算法在sceneflow上的精度是训练中的最优精度么?我在一张1080Ti上训练了basic模型,batch size = 4,其他参数与文中一致,在训练15个epoch后得到了比paper中更高的精度。此外,我在KITTI2015上进行refine时,发现最佳精度并不是出现在在第300个epoch。不过我的batch size是4,而非源码中的12,我想可能这也会有影响吧?

youmi-zym commented 6 years ago

你好,请问你提交到KITTI榜评估过吗?我下载了作者的代码和模型,跑了一下KITTI-2015的testing and training,发现效果并没有达到作者的2.32。。。具体的可以帮忙看一下我刚刚提交的问题,谢谢!

IceTTTb commented 6 years ago

@youmi-zym 您好,我目前还没有在KITTI上提交结果,但是我也没能在KITTI2015验证集上达到论文中的精度(论文table2中,basic模型是2.09%,stackhourglass模型是1.83%,我的误差是2.17%),我在想这是否和batch size的设置有关。

Sarah20187 commented 5 years ago

@IceTTTb 我用了单块gpu train也达到了作者reported的精度,应该和总的batch size没有关系,只和单块gpu的batch size有关

benzhangwang commented 5 years ago

@youmi-zym 你好,关于你在kitti2015用作者训练好的模型提交后的d1-all为2.92%,而不是作者列出的2.32%,我提交后出现的结果跟你相似,请问你找到原因了么?

youmi-zym commented 5 years ago

不要用kitti 官方的matlab处理,submission跑完之后就直接提交结果

benzhangwang commented 5 years ago

恩,我将pytorch0.4 中的upsample函数的align_corners=True 就可以了,但用作者给出的验证集的d1-all错误率为0.915,也不知道是什么原因

在 2018-11-15 15:20:11,"youmi" notifications@github.com 写道: 不要用kitti 官方的matlab处理,submission跑完之后就直接提交结果

benzhangwang notifications@github.com 于2018年11月15日周四 下午1:11写道:

@youmi-zym https://github.com/youmi-zym 你好,关于你在kitti2015用作者训练好的模型提交后的d1-all为2.92%,而不是作者列出的2.32%,我提交后出现的结果跟你相似,请问你找到原因了么?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/JiaRenChang/PSMNet/issues/87#issuecomment-438919962, or mute the thread https://github.com/notifications/unsubscribe-auth/AQ6nQdT3WeYWu58SD5bdLy5ydjD9CGrLks5uvPeLgaJpZM4XzOO_ .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

youmi-zym commented 5 years ago

嗯嗯,这个也是一个问题,的确需要把align_corners=True。不过0.915是啥??在哪

benzhangwang commented 5 years ago

0.915是我用作者给出的训练好的模型在KITTI2015上的验证集的d1-all错误率,可以私聊哈 我的邮箱benzhang_wang@163.com,有问题还的继续向你请教。

youmi-zym commented 5 years ago

@benzhangwang OK, is that means the 3px error arriving at 0.915% on 40 validation images of KITTI Training data ?

benzhangwang commented 5 years ago

是的,可能是文章中列出夫人1.83% 3px error是训练了300 epochs ,而作者给出的训练好的模型应该是训练了1000 epochs,其验证集3px error 是0.915, kitti 2015 leaderboard上 测试集D1-all 是2.32%, 那这两者之间相差还挺大的哈。

在 2018-11-16 15:25:32,"youmi" notifications@github.com 写道:

@benzhangwang OK, is that means the 3px error arriving at 0.915% on 40 validation images of KITTI Training data ?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

JiaRenChang commented 5 years ago

Hi, all The published model is trained for 1000 epochs using ALL KITTI 2015 training set. This model is for our KITTI submission. It is normal that the 3-px error of val set is low because we used all training set.

chujianjiaqi commented 5 years ago

你好,请问一下,你得到的3px error都是使用finetune在KITTI的训练集上得到的吗,我这里不是很懂,用submission进行测试不是只能得到视差图,并没有所谓的3px error

IceTTTb commented 5 years ago

@chujianjiaqi 对于训练集,你可以用kitti官方给的matlab的代码去计算3px error;对于测试集,你可以提交到kitti上去看最终结果的精度。

chujianjiaqi commented 5 years ago

对于训练集不是在进行finetune的时候就直接有3px error吗