Datasets:

ArXiv:
License:

Track 1 Classes

#13
by Yousof10 - opened

This is my first time joining the AI City Challenge, and I noticed that some classes, like FourierGR1T2→4 and AgilityDigit→5, only appear in the test set, not the training set (though I might have missed them in the training data).

The FAQ states: “Additional manual annotations on our testing data are strictly prohibited.”

Given that, is it valid to use models like CLIP or Grounding DINO to handle these unseen classes? Can I also assume that classes like 4 and 5 have consistent color patterns and use a color-based classification approach?
Is there a known method that participants typically rely on each year for this?

Thank you 😄

NVIDIA org

@Yousof10 Thanks for your questions. The FourierGR1T2 and AgilityDigit object classes exist in the training scenes such as Warehouse_014 and validation scenes such as Warehouse_016. You are also allowed to use public pre-trained models like CLIP and Grounding DINO.

zhengthomastang changed discussion status to closed

Sign up or log in to comment