search

open-mmlab / mmdetection3d

OpenMMLab's next-generation platform for general 3D object detection.
https://mmdetection3d.readthedocs.io/en/latest/
Apache License 2.0
5.21k stars 1.53k forks source link

Support ONCE dataset #1353

Open ZCMax opened 2 years ago

ZCMax commented 2 years ago

This is a development requirement documentation of a feature we wish to collaborate with the community. In this issue, we wish to welcome the community developers to participate in the support of ONCE dataset.

Steps

Resources

zehuichen123 commented 2 years ago

Hi, I am interested in this work. However, I can not align the performance on CenterPoint. Could you provide some advice? Perhaps about whether I should change the yaw angle to align with MMDetection3D? Since there is no explanation on how the direction is designed in ONCE dataset, I am quite confused on its \theta value.

ZCMax commented 1 year ago

We suggest you to transform the point clouds coordinates and annotations from ONCE coordinate systems to our LiDAR coordinate system for training and inference. About the visualization problem, I think it's due to the relative position for two coordinates is the same.

loveSnowBest @.***> 于2022年6月14日周二 14:24写道：

Hi, I am currently working on the ONCE dataset and trying to support it on mmdet3d. However, I find that the XYZ coordinate system of ONCE is different from the one used in mmdet3d. In ONCE, the y-axis is towards the back and the x-axis towards the left, while in mmdet3d LiDARInstanceBox, the y-axis is towards the left and the x-axis is towards the front. But interestingly, when I put the original annotation from ONCE to the mmdet3d and visualize them with Open3D, it seems there is no XYZ shift. What is the possible reason for this? I suppose the point cloud coordinate system is also in the ONCE-XYZ system, so the relative position is right. Is that right? Besides, If I want to run the ONCE dataset with mmdet3d models, should I transform the annotations into the mmdet3d-style XYZ coordinate system? (as well as the point .bin files) [image: Uploading wecom-temp-45d0d55b390d49e5f097a752c0e88afd.png…]

— Reply to this email directly, view it on GitHub https://github.com/open-mmlab/mmdetection3d/issues/1353#issuecomment-1154766572, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIZ5DABGGIWHWCOE6HJJQ2TVPAQRTANCNFSM5SCI5VHA . You are receiving this because you authored the thread.Message ID: @.***>

AvivSham commented 1 year ago

Are there any updates regarding this issue? I would like to train some of the models here using ONCE dataset. Do you support this dataset? if not can you please explain what are the differences and whats need to be changed to make it work?

OrangeSodahub commented 1 year ago

@ZCMax Hello all, I've drafted a pr #2257 to support ONCE dataset, basically it now support train/eval/test pipeline. However there must exist some questions about coordinate system, maybe I need your help with this little problem!