resource not found

[robo-alex]

Hi! Thanks for your contributions!
I met some problems while downloading the dataset.
I was trying to download the replica dataset using this command

omnitools.download all --components replica --subset fullplus \
  --dest ./omnidata_starter_dataset/ \
  --connections_total 40 --agree

However, it shows an error indicating resource not found

Would you please help me with that?
Thanks in advance!

[roman-bachmann]

Hi!
We are currently in the process or restructuring the files on the volume. We'll update this issue when the new download scripts are available.
Best, Roman

[ykarmesh]

Hi @roman-bachmann, does that mean that currently no data is available for download? Also can you give a timeframe. when you are thinking of making this data available?

[robo-alex]

Hi! We are currently in the process or restructuring the files on the volume. We'll update this issue when the new download scripts are available. Best, Roman

That's Great! I really appreciate it!

[alexsax]

@mtzhang1999 @ykarmesh the restructuring should be complete now, and I just pushed a new version of the tool that works with the new remote file structure. Please make sure to upgrade with pip install --upgrade omnidata-tools.

Also, note on the data: some of the labels for HM3D are not yet complete, but RGB, depth, and normals are ready and available now. We also plan to release instance segmentation labels for HM3D, contingent on the HM3D authors making them available for the meshes. Lastly, we'll probably regenerate hypersim-derived labels once this issue is completely resolved.

[mikeroberts3000]

Hi friends, the issue with Hypersim camera parameters has been resolved.

[mikeroberts3000]

Hi friends, I'm just double-checking, did you ever figure out how to generate the appropriate Hypersim metadata in your Blender rendering setup?

[alexsax]

Thanks for checking! Not yet.

I think the discrepancy is that Blender doesn't allow lens tilt but does have a 'shift' parameter

(image stolen from here).

There must be a way to handle this, but I haven't cracked it yet

[mikeroberts3000]

Hi @alexsax, thanks for the update. I have two silly questions.

• Why is it important to render the Hypersim scenes with Blender? What exactly is the metadata you want to render? As far as I'm aware, you're not attempting to render the scenes photorealistically. In this case, it might be easier to generate the desired metadata directly in V-Ray using a custom rendering pass. V-Ray has a ton of options for rendering different types of ground truth, so you would not need to mess around with any weird camera parameters, and it would be computationally cheap because V-Ray wouldn't need to do any accurate lighting and shading calculations. Alternatively, it might be easy to write a few extra lines of code in the OpenGL renderer that ships with Hypersim to generate all the metadata you need. To summarize, is there anything that Blender is buying you here, other than convenience and compatibility with the rest of the Omnidata pipeline?
• Does Blender have any way of specifying custom camera projections in the form of a 2D texture? This is a useful feature when doing special effects for movies, so it's possible that it is already implemented in Blender. V-Ray has this feature. It's handy because it gives you complete control over the outbound camera ray at every pixel. If Blender has this functionality, we could just define a 2D texture for each Hypersim scene with all of the appropriate camera rays. This is mostly already implemented in the Hypersim repo, we would just need to save the rays to a 2D texture (which could vary per scene).

I'd be happy to help out with either of these efforts, so I'm trying to better understand what you're trying to do.