search

TRAILab / CaDDN

Categorical Depth Distribution Network for Monocular 3D Object Detection (CVPR 2021 Oral)
Apache License 2.0
366 stars 62 forks source link

result on waymo dataset #29

Closed walterchenchn closed 3 years ago

walterchenchn commented 3 years ago

Thanks for your contribution!

Waymo dataset is quite large and more diverse. Why didn't you use a look-around camera for detection on the Waymo？

codyreading commented 3 years ago

Hi Walter!

We didn't do the full 360 degree camera rig on Waymo due to GPU memory constraints on processing multiple images simultaneously. This would require us to reduce our model capacity, greatly impacting results.

walterchenchn commented 3 years ago

Can you publish the model config of waymo dataset?

codyreading commented 3 years ago

Hi Walter,

I don't have a config for the Waymo dataset for the updated version released here. The config should be the similar to the KITTI config, but point cloud range should be changed to POINT_CLOUD_RANGE: [2, -25.6, -4.0, 55.76, 25.6, 4.0]. Additionally, the Waymo images were down-sampled to meet memory constraints. The current version doesn't have support for Waymo images in the data loader, so this will have to be implemented as well.

Hope this helps!

vobecant commented 3 years ago

Hi @codyreading,

do you intend to release the config file for the Waymo dataset?

Thank you very much!

codyreading commented 3 years ago

Hi and thanks for the interest!

Yes, I plan to release the config, however this also requires adding support for Image Data Loading from the Waymo Dataset.

vobecant commented 3 years ago

Hi @codyreading, thank you very much for your response, I am looking forward to the release. Have you got any idea of when it might be? Regarding the support of data loading: don't you already have this in pcdet/datasets/waymo/waymo_dataset.py? If not, would you please release this loader along with the configs?

Thanks!

codyreading commented 3 years ago

Hello!

The functionality in pcdet/datasets/waymo/waymo_dataset.py doesn't currently support images, only LiDAR point clouds. The code that I used for the Waymo dataset experiments in the paper is outdated, so this would not be able to be used here.

In terms of timeline, I'm not sure as this is a lower priority item for me at the moment. If you need this urgently I would recommend implementing this yourself.

walterchenchn commented 3 years ago

Hi @codyreading I hope to use the look-around camera data of the waymo dataset. But the picture frame of camera on the side is different with the other. How do you deal with these differences?

codyreading commented 3 years ago

Hi @walterchenchn,

Again, we didn't use the other camera and only used the front camera for our experiments. If you would like to use multiple cameras, you could try using one network for all the cameras, but you might have to have separate networks for each camera.

walterchenchn commented 3 years ago

thanks!

vobecant commented 3 years ago

Hello!

The functionality in pcdet/datasets/waymo/waymo_dataset.py doesn't currently support images, only LiDAR point clouds. The code that I used for the Waymo dataset experiments in the paper is outdated, so this would not be able to be used here.

In terms of timeline, I'm not sure as this is a lower priority item for me at the moment. If you need this urgently I would recommend implementing this yourself.

Hi @codyreading, thank you for your reply. Do you think that you could please share with me at least the outdated version of the dataloader? I might be able to convert it to a usable version. Thanks!

codyreading commented 3 years ago

The outdated version of the data loader is essentially the KITTI data loader, as I converted the Waymo Dataset to a KITTI format for use in my network. If you plan to start there, I would recommend using the current KITTI dataloader.