search

weiyithu / NerfingMVS

[ICCV 2021 Oral] NerfingMVS: Guided Optimization of Neural Radiance Fields for Indoor Multi-view Stereo
MIT License
434 stars 58 forks source link

Why sparse depth map resolution 288x364 ? #29

Open Gardlin opened 2 years ago

Gardlin commented 2 years ago

Hi,

I try to generate a larger resolution (eg. 480x640) depth prior map, however ,the visible results seems get worse the 288x364 resolution. Have you tried larger reslution depth prior or that's why you use sparse resolution depth prior.

Another question in your paper, I see the comparison results on the novel views , I'm wondering what's the exact novel view ,does it mean the the testing images or the rerendered views(eg. 120 poses) .Because I don't find the code of rendering the 120 poses images in your repo , which needs depth_prior and depth_confidence. image

weiyithu commented 2 years ago

The resolution 288x364 is same with the setting in CVD. In our extension work, we find that for most of scenes high resolution will bring higher performance. However, for a few scenes it will degrade the performance but we have not found the reasons. We predefine the novel views of each scene in 'test.txt', which are evenly selected from all views.