About how to exclude the first two frames in the evaluation protocol

[flotaas]

Thank you for your work!
I noticed that "PARA-Drive: Parallelized Architecture for Real-time Autonomous Driving" mentioned "State-of-the-art end-to-end AV stacks use temporal information in inputs, and as a result have inferior performance in the first frame due to zero initialization of input features and ego vehicle’s state. AD-MLP addresses this by excluding the first two frames in their evaluation protocol".
I wonder that in the code, how AD-MLP achieve this, i.e., excluding the first two frames in the evaluation protocol?

[scok30]

Hello,
Thank you for your interest in our work. Regarding the description you mentioned about referencing our work in later tasks, I would like to clarify that there is no causal relationship here. We did not exclude earlier frames from the metric due to their poor performance. Instead, we followed the testing methods used in prior work. In fact, earlier future frames are generally easier to predict.
As for the evaluation code, in addition to the version we provided, you can also look at methods like UniAD, which we have referenced before. The implementation is essentially the same.

[flotaas]

Thank you for your kind reply