search

YoYo000 / MVSNet

MVSNet (ECCV2018) & R-MVSNet (CVPR2019)
MIT License
1.38k stars 303 forks source link

How to reproduce preprocessed data #104

Open demonleach opened 3 years ago

demonleach commented 3 years ago

Hi,

Thank you for sharing codes!

I want to produce the same data that you provide. However, detailed settings for tanks and temples are missing in your MVSNet paper.

  1. When using SfM, did you used original images or images of size (1920, 1056)?
  2. Which algorithm did you use, incremental(sequential) SfM or global SfM?
  3. You used OpenMVG to recover point cloud and camera positions. How did you determinate camera intrinsics?
demonleach commented 3 years ago

@YoYo000

I would like to apply openMVG to my own dataset with the same setup as in the MVSNet paper. Your prompt reply is highly appreciated.

YoYo000 commented 3 years ago

Hi @demonleach we are using a modified incremental SfM pipeline. I doubt that if you are using the raw OpenMVG software you cannot get high-quality results similar to our provided ones.

In our experience, COLMAP SfM is quite stable. And I would recommend you to compare it with OpenMVG for your own dataset.

demonleach commented 3 years ago

Hi @YoYo000

Thank you for your help. When I applied COLMAP to my own dataset, I can get higher quality results than OpenMVG.

By the way, how do you determinate camera intrinsic parameters? I used the automatic reconstruction tool in a variety of scenes, but the camera parameters I got were quite different, even though I was using a single camera model.