search

Lilac-Lee / Neural_Scene_Flow_Prior

Neural Scene Flow Prior (NeurIPS 2021 spotlight)
https://lilac-lee.github.io/Neural_Scene_Flow_Prior
MIT License
120 stars 11 forks source link

Regarding nuScenes dataset used in this paper #7

Closed MinkyoungCho closed 2 years ago

MinkyoungCho commented 2 years ago

Hello! I have been really impressed by your amazing work; thanks for your contributions!

I am now curious about how the nuScenes dataset was modified from the original one here. Especially, when I downloaded the original dataset (Trainval, not mini), the PCD format is .pcd.bin but yours is .npz. I saw below in the paper but I cannot get fully; can I know the exact process you did to the original dataset?

We followed the data processing method in [45] to collect pseudo-ground-truth scene flow.

Thanks in advance!

Lilac-Lee commented 2 years ago

Hi, thanks for the suggestion. We have updated the code and README. You can find the preprocessing code in preprocess_sf_argoverse.py and preprocess_sf_nuscenes.py. Cheers!

MinkyoungCho commented 2 years ago

I sincerely thank you for sharing the code I asked for and your amazing work! Wish you luck.

yanconglin commented 1 year ago

hi @Lilac-Lee, Thank you for releasing the code. Unfortunately, I am not able to re-generate the same nuscenes scene flow dataset using the provided script. There are only 150 generated samples vs 310 in the given dataset (google drive). And the sample indices do not match either. Could you please double check? I would like to re-geneate the dataset because I need the ground plane which has been removed from the given dataset. Many thanks!

Lilac-Lee commented 1 year ago

Hi, we used validation dataset provided in the nuScenes trainval-1.0 dataset, and there are 310 samples.