After modifying the splits/eigen_zhou/train_files.txt, I reproduced the results of the paper

[XIAN-XIAN-X]

Thanks to the authors for their outstanding contribution. I have successfully reproduced the results of the paper. However, I have encountered new confusion.

At first, I failed to reproduce the results of the paper. However, I discovered from the authors' comments that using the previous version could reproduce the results.

So I obtained the previous version by executing the git checkout 6a1e997f97caef8de080bb2873f71cfbad9a8047 command.

Here are the experimental results of my ResNet50 model at a resolution of 192x640.

Compare to the new version,the previous version that affects the experimental results is the 'train_files.txt' file in 'eigen_zhou', which contains about 71k images.However, the training set mentioned in the author's paper contains only 26k images.

In addition, I trained using the previous train split (71k images)on Monodepth2 and achieved results similar to the SQLdepth.

I am very confused by the experimental results.My question is:

1. Did author use a train set containing 71k images during training.
2. How to interpret Monodepth2 achieving similar results when trained on a dataset containing 71k images. @hisfog "Looking forward to your reply.Thanks a lot!

[XIAN-XIAN-X]

您好，我想请问您在复现的时候有没有出现 Hugging Face 的模型库 (huggingface.co) 下载模型文件时出现连接超时问题呢 虽然我怀疑是服务器没办法访问该网站，但是我已经手动下载相关权重放置/.cache/torch/hub/checkpoints/还是报错

你好，我没有出现这样的问题。

[Ecalpal]

Very interesting, thank you for your work.

I have briefly checked it and it is consistent with what you said.

[wangjiyuan9]

Have you tried using other depth estimation models to train on this particular partition?

[BlueEg]

Have you tried using other depth estimation models to train on this particular partition?

The Author of this issue trained the Monodepth2(Resnet34) on this particular dataset, you can check.

[hwlf]

Thanks to the authors for their outstanding contribution. I have successfully reproduced the results of the paper. However, I have encountered new confusion.

At first, I failed to reproduce the results of the paper. However, I discovered from the authors' comments that using the previous version could reproduce the results.

So I obtained the previous version by executing the git checkout 6a1e997 command.

Here are the experimental results of my ResNet50 model at a resolution of 192x640.

Compare to the new version,the previous version that affects the experimental results is the 'train_files.txt' file in 'eigen_zhou', which contains about 71k images.However, the training set mentioned in the author's paper contains only 26k images.

In addition, I trained using the previous train split (71k images)on Monodepth2 and achieved results similar to the SQLdepth.

I am very confused by the experimental results.My question is:

1. Did author use a train set containing 71k images during training.
2. How to interpret Monodepth2 achieving similar results when trained on a dataset containing 71k images. @hisfog "Looking forward to your reply.Thanks a lot!

I also want to know about this issue，Have you cleared up your doubts？

[XIAN-XIAN-X]

Thanks to the authors for their outstanding contribution. I have successfully reproduced the results of the paper. However, I have encountered new confusion. At first, I failed to reproduce the results of the paper. However, I discovered from the authors' comments that using the previous version could reproduce the results. So I obtained the previous version by executing the git checkout 6a1e997 command. Here are the experimental results of my ResNet50 model at a resolution of 192x640. Compare to the new version,the previous version that affects the experimental results is the 'train_files.txt' file in 'eigen_zhou', which contains about 71k images.However, the training set mentioned in the author's paper contains only 26k images. In addition, I trained using the previous train split (71k images)on Monodepth2 and achieved results similar to the SQLdepth. I am very confused by the experimental results.My question is:

1. Did author use a train set containing 71k images during training.
2. How to interpret Monodepth2 achieving similar results when trained on a dataset containing 71k images. @hisfog "Looking forward to your reply.Thanks a lot!

I also want to know about this issue，Have you cleared up your doubts？

Not yet.