chaytonmin / Occupancy-MAE

Official implementation of our TIV'23 paper: Occupancy-MAE: Self-supervised Pre-training Large-scale LiDAR Point Clouds with Masked Occupancy Autoencoders
Apache License 2.0
252 stars 18 forks source link

Masking Implementation #4

Closed YilmazKadir closed 1 month ago

YilmazKadir commented 2 years ago

In the paper, you say that you mask the point cloud similar to MAE but I could not find the details. In MAE image is divided into patches that contain 14x14 pixels and then randomly masked at patch level. Do you also divide the scenes into 14x14x14 voxels similarly, or do you just randomly select 50% of voxels in each scene and mask them?

chaytonmin commented 2 years ago

just randomly select 50% of voxels in each scene and mask them

We just randomly select 50% of voxels in each scene and mask them. The pathes in MAE are similar to the voxels in our Voxel-MAE.

alpacaduby commented 1 year ago

The paper indicates that mask ratio varies with distance, but I can't find the related code.Can you please tell me where it is,or you just set the mask ratio to 50% in spite of distance?

chaytonmin commented 1 year ago

The paper indicates that mask ratio varies with distance, but I can't find the related code.Can you please tell me where it is,or you just set the mask ratio to 50% in spite of distance?

Please refer to:

https://github.com/chaytonmin/Voxel-MAE/blob/459fafc67300a63a345edeb4701cd2adf24e4838/tools/cfgs/kitti_models/voxel_mae_kitti.yaml#L15

https://github.com/chaytonmin/Voxel-MAE/blob/459fafc67300a63a345edeb4701cd2adf24e4838/pcdet/models/backbones_3d/voxel_mae.py#L157

YilmazKadir commented 1 year ago

But then you mask 50% (0-30) and 30% for (30-inf). In the paper, you claim "For voxels within 0-30 meters, 30-50 meters, and > 50 meters, we set the masking ratio to 90%, 70%, and 50%, respectively."

chaytonmin commented 1 year ago

Voxel-MAE/tools/cfgs/kitti_models/voxel_mae_kitti.yaml

You can set the masking ratio to 90% in https://github.com/chaytonmin/Voxel-MAE/blob/459fafc67300a63a345edeb4701cd2adf24e4838/tools/cfgs/kitti_models/voxel_mae_kitti.yaml#L15