Giters

ZrrSkywalker / I2P-MAE

[CVPR 2023] Learning 3D Representations from 2D Pre-trained Models via Image-to-Point Masked Autoencoders

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

The principle of 2D guided masking

whuhxb opened this issue · comments

commented

Hi @ZrrSkywalker @ZiyuGuo99

Specifically, how to use the 2D saliency map to guide the masking? Also, random masking on the 2D saliency map? Or according to what principle? Thanks a lot.

commented

Thanks for your interest.
We first aggregate multi-view 2D saliency map into a 3D saliency cloud at here. Then, we assign the masking probability of each point by the related saliency value at here. Higher saliency value leads to lower masking probability.

commented

Thanks for your interest. We first aggregate multi-view 2D saliency map into a 3D saliency cloud at here. Then, we assign the masking probability of each point by the related saliency value at here. Higher saliency value leads to lower masking probability.

OK. Thanks a lot. I will try to understand the details again.