docs(how-to-guides): add training docs for centerpoint

[kaancolak]

Description

Prepare documentation about training lidar CenterPoint model

Blocked by:

• refactor(lidar_centerpoint): add training docs autoware.universe#5570
• Add Autoware compatibility mmdetection3d#1

Pre-review checklist for the PR author

The PR author must check the checkboxes below when creating the PR.

• I've confirmed the contribution guidelines.
• The PR follows the pull request guidelines.

In-review checklist for the PR reviewers

The Reviewers must check the checkboxes below before approval.

• The PR follows the pull request guidelines.

Post-review checklist for the PR author

The PR author must check the checkboxes below before merging.

• There are no open discussions or they are tracked via tickets.

After all checkboxes are checked, anyone who has write access can merge the PR.

[kaancolak]

Hello @yukke42 -san, @miursh -san, @yukkysaito -san, @mitsudome-r -san

I prepared an overview section of lidar CenterPoint training under Autoware documentation and training steps under the "lidar_centerpoint" package.

Before we release the fork of the mmdetection3D and sample_dataset(Tier4 Format) under Autoware Foundation, please review the dataset and fork of the mmdetection3D repository. I am waiting for your feedback.

Fork of the mmdetection3D (dev-1.x-autoware branch):
Fork of the mmdetection 3D contains,

• PyTorch to ONNX converter
• Point Encoder optional use_voxel_center_z parameter
• Tier4 dataset class
• Custom metric calculator including NuScenes evaluation metrics (Original repository just support NuScenes dataset)

link : https://github.com/kaancolak/mmdetection3d/tree/dev-1.x-autoware

Sample Dataset(Tier4 Format)

• Contains 600 lidar frames that are 3D labeled in 2 hz(6905 cars, 3951 pedestrians, 75 cyclists, 162 buses, and 326 trucks).

link: https://drive.google.com/drive/folders/15VeeXje1La0FwUyt1hU8pKXXfcCvRyiJ

[mitsudome-r]

@kaancolak FYI, I have created the fork of MMdetection under Autowarefoundation GitHub: https://github.com/autowarefoundation/mmdetection3d.
You should have access to create branch as well so feel free to port your work to the fork.

[miursh]

@kaancolak
I sincerely apologize for the delay in reviewing.
I have noted some comments and change requests regarding the dataset:

1. A smaller file size for the T4 dataset is preferable, as the current size is leading to prolonged download times.
2. The directory structure for each dataset should adhere to the guidelines outlined in this document. Specifically, there should be an 'annotation/data/input_bag' directory directly under 'T4_dataset_ID' (or 'scene_name').
3. If the annotation frequency (Hz) is lower than that of the data, frames that were not annotated should not be included in sample.json. Additionally, for those frames, the is_key_frame in sample_data.json should be set to false. This might be due to our tools bug.
4. It appears that the calibration, especially between CAM_BACK and base_link, is not well-aligned. Is this OK?

[kaancolak]

Hi @miursh -san, thank you for your feedback. Following that, I plan to filter out non-annotated frames from the dataset. This will reduce the total size from 32 GB to approximately 6 GB and eliminate the need to deal with keyframes.

I will also inspect the calibration of the back camera. If there seems any issue, I can remove them from the dataset.

[kaancolak]

Miura-san (@miursh),

I reloaded the dataset based on your feedback. Can you take a look and give feedback?

The total size decreased by around half of the previous version. (~13 GB)

The back camera is mounted on the rear glass of the vehicle and that part can be opened and closed as a car luggage. Somehow the calibration seems distorted, especially at a far distance. I removed it from the dataset.

Dataset link: https://drive.google.com/drive/folders/1qpts1HlOT3PdvrV_dRd1Q9syY2fhl8pO

[miursh]

@kaancolak I apologize for the delay in responding as I was taking a few days off.
The updated dataset looks good to me! Thank you for the change!

[kaancolak]

Thank you so much @miursh -san. I will ask to Fatih upload the dataset to AWS.

[stale]

This pull request has been automatically marked as stale because it has not had recent activity.

[github-actions]

Documentation URL: https://autowarefoundation.github.io/autoware-documentation/pr-471/
Modified URLs:

• https://autowarefoundation.github.io/autoware-documentation/pr-471/how-to-guides/training-machine-learning-models/training-models/

[stale]

This pull request has been automatically marked as stale because it has not had recent activity.

[kaancolak]

Hi @beginningfan, could you also review this document?