dvlab-research / FocalsConv

Focal Sparse Convolutional Networks for 3D Object Detection (CVPR 2022, Oral)
https://arxiv.org/abs/2204.12463
Apache License 2.0
371 stars 35 forks source link

how to generate 1/4 nuscenes data on openpcdet #24

Closed wenchened closed 2 years ago

wenchened commented 2 years ago

hello, you have give the 1/4 nuscenes data in centerpoint based pkl as [infos_train_mini_1_4_10sweeps_withvelo_filter_True.pkl], so i want to now how to generate 1/4 data based on openpcdet. Thank you very much!

yukang2017 commented 2 years ago

Hi,

Suppose that you already have the full training pkl file of OpenPCDet, nuscenes_infos_10sweeps_train.pkl

You can open it with

train_infos = pickle.load(open('nuscenes_infos_10sweeps_train.pkl', 'rb')) and filter the 'token' of it to match that of tokens in the 1/4 nuscenes data infos I provided.

For example,

filtered_infos = []
for info in train_infos:
    if info['token'] in info_tokens_1_4:
        filtered_infos.append(info)

And then save the filtered_infos. info_tokens_1_4 is the tokens from that of mine.

wenchened commented 2 years ago

Do you me you select some tokens by yourself ? Is the select randomly or follow some rules? Thank you very much!!

yukang2017 commented 2 years ago

I selected it by an interval of 4. https://drive.google.com/file/d/19-Zo8o0OWZYed0UpnOfDqTY5oLXKJV9Q/view?usp=sharing

wenchened commented 2 years ago

Thanks! It really helps!

jyothsna-phd22 commented 5 months ago

Hi, can u pl share the complete dataset training pkl file?