Performance gap between the author's report and your implement #5
Description
Thanks for your great work.
When I read your submit report to ML Reproducibility Challenge 2020, I find one place which is greatly different from the author's results and I also meet the problem when I try to reproduce johnston's (origin author) work.
In the ablation study, as your report, ResNet18+DDV without any other trick can not improve the performance of baseline Monodepth2 (Same as you, My reproduction can not too), even hamper the results. This result is totally different from johnston's and it proves the ddv is actually useless. Furthermore, the uncertain map is also different from his report (I'm sorry to find that you have not reported it too).
As you say, the author has help you reproduce the codes. I want to ask what causes the performance gap between the reproduction and johnston's result? Have he explained? Whether he uses some other tricks to improve the performance? I try to connect to the author but did not get reply.