Learning-Deep-Learning

PillarMotion: Self-Supervised Pillar Motion Learning for Autonomous Driving

August 2021

tl;dr: Self-supervised pillar motion learning.

Overall impression

This paper is benchmarked against MotionNet. The backbone follows that of MotionNet, but instead of using bbox as annotation, it uses the optical flow information from camera image as supervision signal.

Pillar Motion can be used for better tracking, and planning in case of an unknown object.

Key ideas

Technical details

Notes