Any plan to support mode datasets such as nuScenes?

[lucasjinreal]

Waymo and nuScenes are really good dataset to test.

[sshaoshuai]

I will find time to support Waymo dataset recently, and contributions are welcomed to support nuScenes.

[lucasjinreal]

that sounds very nice to support waymo.

Hoping for your workaround!

[sharma-n]

Hey @sshaoshuai thanks for sharing your work! I would like to take a crack at training using the nuScenes dataset. I am planning to use the nuscenes_to_kitti code to rewrite the nuscenes dataset into the KITTI format. Just wanted your feedback on what are the things I will need to refactor. Here are some things I can see I will need to do:

1. Export the nuScenes dataset intothe KITTI format uisng export_kitti.py
2. Refactor pcdet\datasets\kitti\kitti_dataset.py with the correct filenames, etc.
3. Edit the .yaml config file to train new network.

I have a couple of quick questions:

1. Given that nuScenes detection classes are different from KITTI, will changing the .yaml file be sufficient to retrain, for example, PartA2?
2. The ANCHOR_GENERATOR in PartA2 is defined for each class separately. How did you decide what values work well for each class? For example, for Car we have anchor_range: [0, -40, -1.78, 70.4, 40, -1.78] while for Pedestrian we have anchor_range: [0, -40, -0.6, 70.4, 40, -0.6]. what do these values mean? Any way to calculate them for the different classes in the nuScenes dataset?

Cheers!

[Nuri-benbarka]

Hi @jinfagang,
There is this repo Waymo_Kitti_Adapter which converts the Waymo dataset into Kitti format. This might be a fast solution for you.
I hope you will find it useful.

[yukkysaito]

Hi @sshaoshuai
It looks like it is registered in the waymo leaderboard. Is there any update?
https://waymo.com/open/challenges/entry/?challenge=DETECTION_3D&email=shaoshuaics@gmail.com&timestamp=1590993748503967

[rvalienter90]

Hi
Also for another dataset, this repo https://github.com/yzhou377/argoverse-kitti-adapter converts the Argoverse dataset into KITTI format. With good results when using P-RCNN. I hope you will find it useful.

[sshaoshuai]

Hi all, the NuScenes dataset has been supported with two strong baselines.