search

hustvl / VAD

[ICCV 2023] VAD: Vectorized Scene Representation for Efficient Autonomous Driving
https://arxiv.org/abs/2303.12077
Apache License 2.0
641 stars 63 forks source link

Perception without tracking task #36

Open chenzyhust opened 10 months ago

chenzyhust commented 10 months ago

Why does Agent Query directly perform motion transformer without performing tracking tasks? Will this reduce performance?

rb93dett commented 9 months ago

Yes, VAD does not explicitly incorporate tracking. We found that using implicit features can actually achieve pretty good motion prediction results, and VAD primarily focuses on planning. Theoretically, incorporating tracking could further improve motion performance, but it requires a trade-off in terms of computational complexity and speed.