Non-key_frame radar sample_data labeling

#1

Hi,

The sample/sample_annotation/key_frame is 2Hz in frequency. It means that the furthest sample_data would be about (1000ms / 2Hz / 2 = 250ms) away from the associated key_frame.
In this period, the car may move (250ms * 80MPH ~= 8.9meters), if the ego car is moving in +40MPH and the other car is moving in -40MPH.
In this case, how can we label the radar points in non-key_frame sample_data, considering that the true radar points may be 8.9meters away from the bounding box in the key_frame?

Thanks,
Feng

#2

Hi @fjin4,

  • I am not sure what you want to do. If you want only accurate annotations, then you should only use the sample_data that correspond directly to a sample (and its annotations).
  • If you need more training data and can accept a slightly lower quality, you can do linear interpolation, just as we do in https://github.com/nutonomy/nuscenes-devkit/blob/master/python-sdk/nuscenes/nuscenes.py#L301 . In my experience this still works well enough.
  • I think the vehicles rarely move faster than 10m/s.
#3

Hi,

What I want is to get the bounding box in the non_key_frame sample_data, and then to label the radar points in that sample_data.
And the interpolation looks good to me for low speed case.

Thanks,
Feng

#4

Yes, then I think you should use the provided method.

#5
#6
#7

It’s not clear to me how I get a handle to the non key-frame data. If I get the first_sample_token, the next token is another key-frame.

***NVM, figured it out, you just get the sensor sample data token from the first sample, then use the next field from the sample data to get the next sweep

1 Like
Radar data labels