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

Dataset without train part #5

Closed hi-zhengcheng closed 2 years ago

hi-zhengcheng commented 2 years ago

Hi @Lilac-Lee: interesting paper.

In paper's table 1, it says that

nuScenes Scene Flow Train: 1,513 samples, Test: 310 samples

Argoverse Scene Flow Train: 2,691 samples, Test: 212 samples

Following the download link given in this repository, it seems that only val part is included in these two datasets. Could you also provide the download link for the train part or some instructions to download this kind of data? It will be appreciated if the train part is also included.

Lilac-Lee commented 2 years ago

Hi, our method did not use any training data. And these real-world autonomous driving scenes do not have ground truth for scene flow. Table 1 only gives an idea of the original dataset split. Cheers.

Szy-Young commented 2 years ago

Hi @Lilac-Lee: First thanks for sharing the wonderful work.

For nuScenes & Argoverse Scene Flow dataset, I see you follow [45] to collect pseudo-ground-truth scene flow. Since [45] has not released the code, will you consider sharing your implementation of data processing? This would be very helpful for future learning-based methods to follow your work and compare with it on these two benchmarks.