BEV grid resolution

[tjqansthd]

Hi! thanks for the great work.

If I want to train BEVDepth using 256x256 BEV grid resolution, How should I modifiy the setting? Do I just have to change

1. 'voxel_size=[0.2, 0.2, 8]' to 'voxel_size=[0.1, 0.1, 8]' 
2. 'x_bound: [-51.2, 51.2, 0.8]' to 'x_bound: [-51.2, 51.2, 0.4]'
3. 'y_bound: [-51.2, 51.2, 0.8]' to 'y_bound: [-51.2, 51.2, 0.4]'

in base_exp.py, or is there anything else to consider?

Thanks.

[yinchimaoliang]

Hi there, you also need to consider out_size_factor.

[tjqansthd]

Oh, I see. It is used in calculating feature_map_size. Thank you!

[yukaizhou]

Oh, I see. It is used in calculating feature_map_size. Thank you!

Hi there。After following this modification, is your result normal? According to your description, the same modification was made in the original code. specifically is backbone_conf： 1，x_bound: [-51.2, 51.2, 0.8]-->:[-51.2, 51.2, 0.4]; 2, y_bound: [-51.2, 51.2, 0.8]-->:[-51.2, 51.2, 0.4]; bbox_coder： 1， out_size_factor=4 --> 2, 2, voxel_size=[0.2, 0.2, 8] --> [0.1, 0.1, 8]; train_cfg： 1， out_size_factor=4 --> 2, 2, voxel_size=[0.2, 0.2, 8] --> [0.1, 0.1, 8]; test_cfg： 1， out_size_factor=4 --> 2, 2, voxel_size=[0.2, 0.2, 8] --> [0.1, 0.1, 8]; Other than that there are no changes. But there is no result at all, that is, mAP=0.

[tjqansthd]

Oh, I see. It is used in calculating feature_map_size. Thank you!

Hi there。After following this modification, is your result normal? According to your description, the same modification was made in the original code. specifically is backbone_conf： 1，x_bound: [-51.2, 51.2, 0.8]-->:[-51.2, 51.2, 0.4]; 2, y_bound: [-51.2, 51.2, 0.8]-->:[-51.2, 51.2, 0.4]; bbox_coder： 1， out_size_factor=4 --> 2, 2, voxel_size=[0.2, 0.2, 8] --> [0.1, 0.1, 8]; train_cfg： 1， out_size_factor=4 --> 2, 2, voxel_size=[0.2, 0.2, 8] --> [0.1, 0.1, 8]; test_cfg： 1， out_size_factor=4 --> 2, 2, voxel_size=[0.2, 0.2, 8] --> [0.1, 0.1, 8]; Other than that there are no changes. But there is no result at all, that is, mAP=0.

Hi, if you change voxel_size=[0.2, 0.2, 8]-->[0.1, 0.1, 8] and grid_size=[512, 512, 1]-->[1024, 1024,1] in train_cfg, test_cfg, bbox_coder, you don't need to change out_size_factor=4 --> 2.