Will pretrained weights be available?

[zjh21]

Thank you for your great work.

As I see from your to do list, you plan to release inference script for pretrained models. Does it mean the pretrained weights on your dataset will be made available?

[eisneim]

+1

[yashkant]

Hey folks, we won’t be releasing the pretrained weights, as we haven’t got the legal approval for it (so far).

The inference script would be for models trained on sample ava-256 data provided in the repo.

Apologies for the confusion!

[eisneim]

@yashkant thanks for your reply this is a great project! thanks for providing training code

[yashkant]

By the way, it should be easy to prepare the packaged samples stored in numpy files from raw Ava-256 dataset here; and for similar other datasets (like objaverse and so on).

Also, I will put in a script that takes the raw Ava-256 dataset and recreates the samples (ie, epipolar rays and spatial anchor, etc) in coming days. :)

[yashkant]

Hey folks,

You can refer to this gist which contains helper methods + details how we created Plucker Rays and Spatial Anchors from Ava-256 and Goliath datasets: https://gist.github.com/yashkant/971e205d85b15e17d20d33edd29d6016

I have added comments where new fields such as headpose or 3d keypoints are used from Ava-256 or Goliath raw datasets to generate the anchor.

You can use/refer these methods to create these fields on your own custom datasets! Please feel free to ask any follow-ups if you see anything missing or unclear!

Hope it helps!

Todo: I will add these methods as a standalone script in Pippo repository working with raw Ava-256 / Goliath dataset in future - as I find more time and if the gist is still unclear or complicated.