Low quality
Closed this issue · 2 comments
I'm trying to reproduce the results in the paper, following your instructions, your code and your pre-trained model, but the quality of the depth-maps and fused point clouds are far from what is displayed in the paper. What might be the problem? Did you use a different training procedure? Can you share the model you used for testing and how did you train it?
The visual results on the T&T dataset are trained on the DTU dataset first and fine-tuned on the BlendedMVS dataset. We release the model pre-trained on the DTU dataset which only contains small objects (far from large scenarios on T&T). Since our original model on BlendedMVS is broken, you may fine-tune the model yourself to reproduce the results.
Also, the fusion parameters when fuse depths into point clouds should be tuned. You can check the code in fusion folder.