About Multi-image preprocessing, such as dino, voxelize and 3d vae

[LeeCASC]

I want to confirm the experimental process.
First, input 150 images into the Dinov2 model to obtain the feature (150, 1024, 37, 37).
Then, by voxelization, project it to the plane and then project it back to the voxel space and then average each voxel, will the feature (1024, 64, 64, 64) be obtained? Will averaging reduce the number of channels(1024)? How to filter the obtained features from 64x64x64=262144 to 20,000?
Finally, the pre-trained model 3D VAE is used to process the feature. What is this pre-trained model? What data is used to pretrain 3D VAE?

[JeffreyXiang]

Hi, we provide dataset_toolkits that contains the process of dino feature fusion. VAE is also trained with 500K assets we filtered