Running on nuScenes

[SubMishMar]

Hi, Can you share some more information on running this package on nuscenes data?

Which dataset should I download from here? https://www.nuscenes.org/nuscenes#overview

And what should be the structure of the directory which holds the data?

[SubMishMar]

This picture shows my directory structure and the output of the pre_processing step. What am I doing wrong here?

[qqqGpe]

It seems your dataset is not complete, you should download all the items in trainval.

[qqqGpe]

Hi, Can you share some more information on running this package on nuscenes data?

Which dataset should I download from here? https://www.nuscenes.org/nuscenes#overview

And what should be the structure of the directory which holds the data?

I will attach the structure of the data later on.

[SubMishMar]

Okay thanks! The structure will be of great help.

What to do you mean by downloading all items in trainval?

Currently, I have just downloaded the meta data and the radar sensor data from Part 1.

Do I need to download all sensor data from Part 1 or all sensor data from all Parts or radar sensor data from all Parts?

Please let me know.

[qqqGpe]

Okay thanks! The structure will be of great help.

What to do you mean by downloading all items in trainval?

Currently, I have just downloaded the meta data and the radar sensor data from Part 1.

Do I need to download all sensor data from Part 1 or all sensor data from all Parts or radar sensor data from all Parts?

Please let me know.

You should download radar sensor data from all part

[SubMishMar]

Ok let me try that. Just to be sure, either US or Asia or both?

[qqqGpe]

Ok let me try that. Just to be sure, either US or Asia or both? Either of them. After finishing downloading, you should aggregate all radar data into one folder.

[SubMishMar]

Okay it seems to be running ok after I put all radar data in a folder. However, how can I plot the trajectory to visualize it?

The output.csv seems to give me some error metric but can I plot a 2D/3D plot to visualize the trajectory?

Please let me know. Also, the pre-process step took really long. Is that normal?

[qqqGpe]

Okay it seems to be running ok after I put all radar data in a folder. However, how can I plot the trajectory to visualize it?

The output.csv seems to give me some error metric but can I plot a 2D/3D plot to visualize the trajectory?

Please let me know. Also, the pre-process step took really long. Is that normal?

Since the trajectories of nuScenes dataset are divided into multiple small scenes (each of them last for around 20s), in most cases, there are not many overlaps between two scenes, so visualizing the trajectories of the results may not make sense. That's the reason why we only show the localization errors.

If you really want to see the trajectories, you can modify the code in DCRML.cpp to save the final poses rather than the errors of each loop closing pair, and plot the poses if their timestamps are close enough.

Yes, it take a lot of time to save the data, and the most time-consuming module of the pre-process step is to find the loop closing pairs.

[SubMishMar]

Okay thanks. Let me try to figure this out.

[SubMishMar]

Can you explain what these numbers in the pair_list.csv file mean?

[qqqGpe]

from left to right: scene1_id, scene2_id, sample_id_in_scene1, sample_id_in_scene2

[SubMishMar]

I am basically trying to comprehend what you are doing in this package at a broad level.

So, the loop closure part is to figure out which scenes have samples collected from the same geographical locations?

And once you have found the association between scenes you find out the transformation between samples from different scenes?

[qqqGpe]

That's right