The input resolution seems to have a big impact on the model

[puyiwen]

Hi,I used the input size as 576x448,the model can achieve the result of the paper, but when I changed the input size to 224x224, the model rmse dropped from 0.34 to 0.50. I don't understand why this is, is this normal? Should I retrain the mit_b4 model again?Looking forward to your reply.

[fenfeichen]

Hello, when I was doing the experiment, there was always a gap between the experimental results and the author's results. How did you do this experiment

[puyiwen]

Hello, when I was doing the experiment, there was always a gap between the experimental results and the author's results. How did you do this experiment

I just delete the RandomCrop

[puyiwen]

I think the reason is the A.RandomCrop function.When I use this crop function,I find the model works worse, so I delete A.RandomCrop, the model results are similar to those in the paper.

…

------------------ 原始邮件 ------------------ 发件人: ***@***.***>; 发送时间: 2022年8月29日(星期一) 下午2:41 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [vinvino02/GLPDepth] The input resolution seems to have a big impact on the model (Issue #27) Hello, when I was doing the experiment, there was always a gap between the experimental results and the author's results. How did you do this experiment — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[fenfeichen]

Thank you for your generous answer. It is very helpful to me. Thank you again

[sayakpaul]

Have you tried fine-tuning experiments on any other datasets, such as DIODE?

[sunssssssss]

I recently tried to reproduce this model and found the same situation as you. I was wondering how you analyzed and concluded that random clipping needed to be removed to bridge the gap between the two

[puyiwen]

Have you tried fine-tuning experiments on any other datasets, such as DIODE?

Sorry for my late reply. I have not fine-tuned this model on other datasets. However, I used the GLPDepth model with 224x224 input size to inference my real scene, and I found the visualization to be very nice. This will mean that GLPDepth has a certain generalization.

[puyiwen]

I recently tried to reproduce this model and found the same situation as you. I was wondering how you analyzed and concluded that random clipping needed to be removed to bridge the gap between the two

I am not very clear, because I have done a lot of experiments, and I found that the data enhancement method of crop will reduce the accuracy of my model, whether it is RandomCrop or CenterCrop.