search

zzangjinsun / NLSPN_ECCV20

Park et al., Non-Local Spatial Propagation Network for Depth Completion, ECCV, 2020
MIT License
321 stars 57 forks source link

Questions about supervision. #39

Closed jikerWRN closed 2 years ago

jikerWRN commented 2 years ago

Thanks for your good work and open source code!

When I read the paper and code, I have a question. As far as I know, in the CSPN series of papers, the algorithm first obtains an initial depth image, and then it uses iterative spatial propagation to obtain the final depth image. I found that in these algorithms, such as NLSPN, the initial depth map is not supervised. I wonder if joining the supervision of the initial depth will limit the expressive ability of the network? And I would like to ask if you have done similar experiments or verified the performance of the initial depth map？

Thanks and looking forward to your reply!

zzangjinsun commented 2 years ago

Hi @jikerWRN, Of course, I tried intermediate supervision on the initial depth when I was working on the NLSPN. However, I concluded that the intermediate supervision on the initial depth is too strong, and it weakens the role of the propagation part for further improvement. If you are working on the depth completion, please try and share your experience later!

jikerWRN commented 2 years ago

Ok, I just have such a question, and did not conduct related experiments. If I do related experiments later, I will share it through a new issue. Thanks!