google-research / jax3d

Apache License 2.0
735 stars 96 forks source link

How to reduce the size of the final model for MobileNerf? #115

Open SamuelSJTU opened 2 years ago

SamuelSJTU commented 2 years ago

MobileNerf learn polygon and feature map for object. The result is awesome. Do you know any suggestions on how to reduce the final model size?

czq142857 commented 2 years ago

The simplest way is to reduce the mesh grid size (e.g., 128 -> 64), or to reduce the texture image size (e.g., 17 per quad -> 9 per quad). These are also shown in the ablation studies in the paper.

Unbounded 360-degree scenes are different though. The "skyboxes" modeling surrounding environments take way too much space than the object in the center. I have not found a neat solution yet.

XuHangkun commented 1 year ago

Hi, I compressed the model to 1/7 of the original. Firstly, I convert the features (PNG format) to h264 video. You can split the big feature images into N*N small images and merge them into a video in h264 format. The compression is quite robust with the parameter N, I tried 1, 2, 4, and 8 and all these choices can achieve a similar compression ratio. Then, you can compress the model file (in obj format) with the zip.