search

shawLyu / HR-Depth

[AAAI 2021] HR-Depth : High Resolution Self-Supervised Depth Estimation
MIT License
236 stars 27 forks source link

Experimental results without pretrained #9

Closed YagumaLi closed 3 years ago

YagumaLi commented 3 years ago

Hi,

Thanks for your amazing work! It gives me a lot of inspiration. And have you ever tried training your network without pretrained, and test the trained network at different resolutions? If so, could you please give me some cues to find them?

I will appreciate it if you reply to me. Thanks a lot.

shawLyu commented 3 years ago

Thanks for your attention.

  1. Yes, I had tried to train our network from scratch. The performance of this setting is bad, you can see it below. model_name & abs rel & sq rel & RMSE & log_RMSE & /delta < 1.25 & /delta < 1.25 ^2 & /delta < 1.25 ^3 HR-Depth (640 x 192) | 0.130 | 1.004 | 5.077 | 0.207 | 0.847 | 0.950 | 0.978

  2. The resolution of training set and testing set should be consistent. I think this is due to the kernel size of the convolution.

Hope this answer can help you!

YagumaLi commented 3 years ago

Thanks for your reply. This is what I need. Thanks so much.

And can I use this results as comparison in my paper if it is necessary in the future? Should I just cite your paper or this website?

Thanks a lot!

shawLyu commented 3 years ago

Yes, you can use this result in your paper and I think citing the paper is enough.

If you have any other questions, you can email to me. Thanks!

YagumaLi commented 3 years ago

Thanks a lot!