Open NewMesc opened 3 years ago
It's a depth map refinement. It's described in Sec 3.3 of Fast-MVSNet: Sparse-to-Dense Multi-View Stereo With Learned Propagation and Gauss-Newton Refinement. We find the improvement is marginal so we disable it.
Got it. I am trying to reproduce the result on Tanks and Temples which presented in your paper, but only reahced 57.09 on intermediate leaderboard. And I found your newest submission whcih contained pretrained model and fusion codes. Which dataset was this checkpoint trained on? Could this checkpoint reproduce the T&T result presented in paper?
The model is trained on BlendedMVS. This checkpoint is used by the benchmark result. But the fusion script cannot be released. The provided script is re-implemented by me afterward and I didn't test it on this benchmark. So I cannot guarantee that you can reproduce the same result, but it should still be in the top tier.
Hi, thanks for your nice code and work! After reading the release version of code and your paper, I found GNRefine module not described in paper but was a part of VisMVSNet (although related code was comment out). So, what's the use and the role of GNRefine, looking forward to your answer!