How to Label Real-World Data of Category Using Argoverse 2 Prediction Challenge Techniques?

[SunHaoOne]

Hi,
I am currently using an API to train models for submission to the leaderboard in the Argoverse 2 prediction challenge. In addition to this, I'm interested in applying these techniques to real-world data. I'm seeking clarification on how tracks are labeled within the dataset:

• Track Fragment: These are lower quality tracks that may contain only a few timestamps of observations.

• Unscored Track: These tracks are used for contextual input and are not scored.

• Scored Track: These are high-quality tracks relevant to the Autonomous Vehicle (AV) and are scored in the multi-agent prediction challenge.

• Focal Track: This is the primary track of interest in a given scenario and is scored in the single-agent prediction challenge.
  My main question revolves around the criteria for selecting these labels, especially:

• Does a focal track represent tracks that have a longer duration of data?

• Besides duration, are 'scored tracks' also selected based on their proximity to the AV or other factors?
  I'm considering using a similar methodology to label and utilize real-world data to achieve comparable performance. Thank you for any insights or advice on this approach.

[evercherish]

same question, have you figured it out?