search

TRI-ML / dgp

ML Dataset Governance Policy for Autonomous Vehicle Datasets
https://tri-ml.github.io/dgp/
MIT License
94 stars 62 forks source link

feat: Add utils to associate 2d bbox and 3d bbox #157

Closed jia-sun closed 9 months ago

jia-sun commented 9 months ago

This PR adds functions to associate bounding_box_3d of LIDAR to bounding_box_2d for Camera. One of the use case is to get the distance distribution of objects.

yuta-tsuzuki-woven commented 9 months ago

@jia-sun Could you explain more about the use case of this function and why you would like to put this function in the DGP repo? I just thought if the use case is very specific and not useful for other DGP users, it should not be under the DGP repo...

jia-sun commented 9 months ago

Could you explain more about the use case of this function and why you would like to put this function in the DGP repo?

Those functions can be used to generate ground truth such as distance, yaw, pitch, raw, even 3d bounding for 2d objects (car, pedestrian) by utilizing LIDAR 3D annotations. For ADADS, it requires sensor redundancy. This enables camera ML model to generate equivalent outputs as LIDAR ML models. Personally, I think It is a good place to share those functions to avoid duplicated work of individual developers.

yuta-tsuzuki-woven commented 9 months ago

@jia-sun I added some comments

@chrisochoatri Could you also review this PR?