search

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
251 stars 18 forks source link

Is database-sampler used during pretraining? #25

Closed FrontierBreaker closed 1 year ago

FrontierBreaker commented 1 year ago

I previously think the proposed method is an unsupervised one. However, if gt-sampler is used during pretraining, the proposed method is actually not. Could you tell me whether I get the point? Thanks!

FrontierBreaker commented 1 year ago

@chaytonmin

chaytonmin commented 1 year ago

@chaytonmin

This is just the data process of OpenPCDet.

FrontierBreaker commented 1 year ago

So, you mean database-sampler is used for Voxel-MAE pretraining?

Actually, when I run the pretraining script, I find that DB-sampler is used.

chaytonmin commented 1 year ago

So, you mean database-sampler is used for Voxel-MAE pretraining?

Actually, when I run the pretraining script, I find that DB-sampler is used.

Our occupancy labels are generated by the input voxels, which is not related to the gt boxes. No matter what data sample strategy is used.

FrontierBreaker commented 1 year ago

Got that, thanks!!

FrontierBreaker commented 1 year ago

To the best of knowledge, I think the database sampler introduced in voxelmae somehow indicates that the training data is more than pure lidar&image pairs. Instead, It is the combination of lidar&image and pre-generated object database.

chaytonmin commented 1 year ago

To the best of knowledge, I think the database sampler introduced in voxelmae somehow indicates that the training data is more than pure lidar&image pairs. Instead, It is the combination of lidar&image and pre-generated object database.

Yes, it will include the pre-generated object database. However, it does not include the box information in the occupancy labels.

FrontierBreaker commented 1 year ago

Great thanks to your patient feedbacks! : )