Giters

SamsungLabs / imvoxelnet

[WACV2022] ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

transformation in 'create_nuscenes_monocular_infos'

Jiayi719 opened this issue · comments

commented

Hello, I try to understand the transformation you do on 'tmp_box' in the following code:
https://github.com/saic-vul/imvoxelnet/blob/master/tools/data_converter/nuscenes_monocular_converter.py#L108-L114
But I really couldn't make it. Can you explain the insight of this transformation? Thank you.

commented

Hi @Jiayi719 ,

Looks like this function generates monocular_infos and multi_view_infos. The code you mentioned is only connected with monocular_infos. However it is not used anywhere, we only use nuscenes_multi_view_infos_train.pkl and nuscenes_multi_view_infos_val.pkl in the NuScenesMultiViewDataset.

Does this answer help you?

commented

Then how could we get gt boxes from nuscenes_multi_view_infos_train.pkl and nuscenes_multi_view_infos_val.pkl ?

commented

I'm very sorry for this mess. This is also related to your next issue. Can you please try the original nuscenes_infos_train.pkl from tools/nuscenes_converter.py? Looks like current version of code is ready for loading this file. And we can simply ignore nuscenes_multi_view_infos_train.pkl, nuscenes_monocular_infos_train.pkl and tools/nuscenes_monocular_converter.py.

commented

Yes, you are right. I have generated nuscenes_infos_train.pkl and nuscenes_infos_val.pkl by running the following command:

python tools/create_data.py nuscenes --version v1.0-mini --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

And indeed I succeed to run test on this two pkl files by:

python tools/test.py configs/imvoxelnet/imvoxelnet_nuscenes.py work_dirs/imvoxelnet_nuscenes/20210505_131108.pth --show --show-dir work_dirs/imvoxelnet_nuscenes

But I wonder this process execute multi-view detection or just execute detection on each camera image respectively?

commented

It should be multi-view. You can check that 6 images per scene are loaded here and then processed here.

commented

Yes, you are right. I have generated nuscenes_infos_train.pkl and nuscenes_infos_val.pkl by running the following command:

Removed nuscenes_monocular from README in 3512e89.