Is it possible to download subset with depth maps, and extrinsincs

[PruneTruong]

Hi,

Thanks for your great work. Is it possible to download a subset of categories/sequences which contain accurate camera poses along with the depth maps and rgb? I downloaded the many-view test, but most of the sequences there have no depth map, or the depth mask is wrong. I am mostly interested to train NerF per scene.

Thanks

[shapovalov]

Hi Prune, thanks for your interest in the dataset! For training vanilla NeRF you should not need depth maps – or are you trying some variant that uses pre-computed geometry? Cameras and depth maps were evaluated automatically, so some may be incorrect. However, for the majority of sequences there are depth map files for at least part of the sequence frames. Can you post some examples of sequences where depth maps are missing so that I can check? You may also find it helpful to check sequence annotations for point_cloud.quality_score to filter for those where the reconstruction is more confident.

[PruneTruong]

Thanks a lot for your very quick answer! Yes, so I mostly use the depth maps for evaluation, to evaluate the learnt geometry, when training on a few images, but also potentially for debugging purposes. For example, baseballglove, 601_92967_186813 baseballbat, 569_82570_162821 banana, 610_96665_193716
have a fully black depth. There are quite a few sequences in the subset that one can download.

Additionally, some sequences have a weird depth mask (like suitcase, 50_2928_8645), which just looks like noise.

Thanks for the tip of the quality score, I will check it out.

[davnov134]

Hello, the depth maps of the manyview-test sequences have been redacted (i.e. they are all black as you noticed) because they have been obtained by running multi-view stereo that used the hidden unseen views. Intrinsics are correct for all views.