Could you give an example of pandaset ?

[cuge1995]

No description provided.

[jacoblambert]

It's not the prettiest approach but personally I created a piece of code to

• Load pandaset sequence by sequence, frame by frame,
• Bring the LiDAR data and annotations into the Pandar64 ego frame
• Normalize LiDAR intensity from 0-1
• Rename some classes to avoid spaces which will break the kitti_utils loading code, e.g. map 'Emergency Vehicle' -> 'Emergency-Vehicle',
• Pick my favorite camera and obtain the necessary tf information
• Save pointcloud, labels, images and calibration file into KITTI format
• Create validation / training split files

Then I created a dataset class following the KITTI template:

• Change loading functions,
• Making sure to remove the camera->lidar frame TF that standard KITTI needs (set identity TF)

Finally make a dataset_config and pandaset_models with modified parameters such as in nuscenes models (full pcd, smaller voxel size). I had decent results with second_multihead and now training pv_rcnn. But I think evaluating on pandaset is not so fair - there are many boxes without any lidar points because of the dataset creation approach. These are not used in training but are used in evaluation, as far as I know. I think I will create a modified labels list with only bounding boxes that have lidar points and are in range, to give fairer evaluation.

Good luck!

[MartinHahner]

@jacoblambert

It's not the prettiest approach but personally I created a piece of code

Even though you say your code is not the prettiest,
I think all of us would be happy if you create a pull request and share that piece of code anyway. 🙂

[xujinchang]

@jacoblambert can u show your code of pandasets ?

[yangnvzi]

@jacoblambert
hi,I am honored to read your comments.
I recently also try to transfer pandar64 format into kitti format
emm...I I've run into some problems about bring the annotations(cuboids and semsegation ) into the Pandar64 ego frame
1)refer to https://github.com/scaleapi/pandaset-devkit,
demision.x ,demision.y,demision.z dimensions of the cuboid based on the world dimensions
how to bring the annotations cuboids into Pandar64 ego frame?
2)semseg only to include index and class_id
how to Distinguish anyone liar points is pandar64's or pandarGT's

[dkoguciuk]

Hi @jacoblambert ,

can you elaborate why do you suggest to bring lidar & labels to egoframe? I'm trying to understand the logic behind it :)

Best,
Daniel

[lea-v]

Hi,
in case it can help, I've started a merge request in order to integrate pandaset support. (#396)
What it contains is:

• training directly on pandaset data (without converting to kitti format).
• using gt_sampling during training
• pick the categories to train on
• select the lidar(s) to load data for
• save pandaset output in the same format as pandaset cuboids annotations
• visual check that pvrcnn trained on pandaset gives coherent results

The main missing parts are:

• the evaluation code : which is made complicated by the fact that there is no "official" evaluation method for pandaset
• addition of trained models to the model zoo

Feel free to build on this when trying to train with pandaset

Best,
Léa

[Leozyc-waseda]

@jacoblambert
Hi,

Can you tell me how you evaluation pandaset?
It would be nice if there was a code.
I have put pandaset through dataloader generated the data infos.

[github-actions]

This issue is stale because it has been open for 30 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.