Giters

exiawsh / StreamPETR

[ICCV 2023] StreamPETR: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

regarding Motion-aware Layer Normalization

RashoAli opened this issue · comments

commented

Dear author,

Thank you for the great work and the very interesting approach. I am trying to understand the "Motion-aware Layer Normalization" part of the model:

  1. Why does the model calculate 'beta' and 'gamma'? What do these parameters represent? And why add and multiply these parameters with the queries afterward?

  2. For pose, time, and velocity: Are these parameters extracted from the detected queries, or are they part of the ground truth (GT)?

commented

@RashoAli beta and gamma are learnable parameters that implicitly containing the motion information.
pose and time are record in the nusc dataset. velocity is from the prediction results.