search

ShuweiShao / AF-SfMLearner

[MedIA2022 & ICRA2021] Self-Supervised Monocular Depth and Ego-Motion Estimation in Endoscopy: Appearance Flow to the Rescue
MIT License
99 stars 15 forks source link

About reproduction of experimental results #20

Closed adam99goat closed 1 year ago

adam99goat commented 1 year ago

Strongly grateful for your excellent work and sorry to bother you for some puzzle. I tried to retrain the network to acquire the same results of SCARED as yours but failed. The validation of end-to-end and stage-wise is shown below. end stage Though the result is worse than what are reported in your paper, the experimental result of Monodepth2 by myself is strangely better. The abs_rel value in your paper is 0.071. image The data preparation is based on OpenCV which leavage the same video encoder and decoder as ffmpeg. So I can't figure out why but only ask you for favour. Look forward to your reply and thank you again!

ShuweiShao commented 1 year ago

Hi, what version is your pytorch, different versions have an impact on the accuracy. Besides, the OFNet also has an impact on the accuray. Please try to load the given weights of OFNet to reproduce the accuracy of DepthNet in the state-wise manner.

adam99goat commented 1 year ago

Excited and thankful for your reply. My pytorch version is 1.8.0 and torchvision is 0.9.0 . And I would like to know your version if convenient. I will try your valuable suggestion and retrain the net in stage-2 with help of provided weights of OFNet. Thank you a lot!

ShuweiShao commented 1 year ago

My PyTorch is 1.2.0 and torchvision is 0.4.0.

1158191255 commented 1 year ago

Excited and thankful for your reply. My pytorch version is 1.8.0 and torchvision is 0.9.0 . And I would like to know your version if convenient. I will try your valuable suggestion and retrain the net in stage-2 with help of provided weights of OFNet. Thank you a lot!

Unfortunately pytorch 1.2.0 does not seem to support 30-series GPU. pytorch 1.2.0 corresponds to CUDA up to 10.0, but 30-series GPU requires CUDA 11.

adam99goat commented 1 year ago

Excited and thankful for your reply. My pytorch version is 1.8.0 and torchvision is 0.9.0 . And I would like to know your version if convenient. I will try your valuable suggestion and retrain the net in stage-2 with help of provided weights of OFNet. Thank you a lot!

Unfortunately pytorch 1.2.0 does not seem to support 30-series GPU. pytorch 1.2.0 corresponds to CUDA up to 10.0, but 30-series GPU requires CUDA 11.

The same trouble as me ... The experiment in the paper was conducted on Titan RTX.

adam99goat commented 1 year ago

My PyTorch is 1.2.0 and torchvision is 0.4.0.

What a pity I think there still exist some problems. Maybe the data preparation should be based on ffmpeg. So could you share related code or command of pre-processing if convenient? Thanks again!

ShuweiShao commented 1 year ago

Hi，I can't find the preprocessing code now since a long time has passed, maybe you can search for the usage of ffmpeg. When you load the given weight of OFNet, what is the accuracy of training DepthNet in a state-wise way?

adam99goat commented 1 year ago

I feel so sorry for my late reply. Actually when I followed the 2-stages custom training with the given weight of OFNet, the result is shown below. image Also PyTorch 1.2.0 and torchvision 0.4.0.