custom data test reconstruction effect is not good

[fanshixiong]

Hello, thank you very much for your open source. I used the mobile phone to collect data and use colmap sparse reconstruction to get the camera parameters, and then use the conversion code provided by yaoyao colmap2mvsnet.py to get network available input. But in the end the reconstruction works poorly. The depth map seems to have a look, but it is wrong, and the point cloud is very small, only a few hundred KB.

I would like to ask:

1. Do you need to change the parameters when using colmap sparse reconstruction (the default parameters I use)
2. There is obviously a problem with the camera parameters converted by colmap2mvsnet (depth range?)

[prstrive]

In our experiments, we use the camera setting of MVSNet and CasMVSNet. I guess the unsatisfactory results on your custom data may lie on the unsatisfactory camera parameters. Maybe you need to clean the point generated by COLMAP first, and then adjust the camera pose and depth range to be compact.