Good work, but why can't we achieve good results after 100000 iterations of training？

[shingkim]

No description provided.

[shingkim]

[hjxwhy]

Me too, I have trained with this code and data download directly, but the result is bad.

[hjxwhy]

@barbararoessle Hi, I have trained several scenes using this code and your data without making any changes, but some val images have floaters (almost every scene has some images with floaters, and even like a mask before the scene). Have you met the same problems? Here is my log https://drive.google.com/drive/folders/1hW4zORQSeAD5tv86hXl1uO48ZPSHDtIt?usp=sharing

[barbararoessle]

Number of optimization iterations: We found that for the room-like scenes longer optimization, 500.000 iterations, improves the results. Since entire rooms with inside-out viewing direction have more surface area, than f.ex. single objects with outside-in viewing direction, I don't find it surprising that they benefit from more optimization iterations.

Regarding the quality of results and floaters: The stetting with very few views and no sensor depth of the scene is under-constraint and difficult, so it is certainly possible that there are floaters in the rendered test views.
The difference to NeRF without depth priors, however, should be large, like in the comparison videos.

[Quyans]

hi

Me too, I have trained with this code and data download directly, but the result is bad.

hi,me too ,I tried directly using the downloaded data and pretrained model. but the psnr even only 17 average. I test ten images ,all results are bad