Learning-Deep-Learning

FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection

August 2021

tl;dr: FCOS baseline of mono3D.

Overall impression

The majority of the single stage mono3D methods since SMOKE all use CenterNet as baseline. This paper switches it to FCOS and achieves good results.

Objects are distributed to different feature levels with the consideration if the 2D scales (from reprojected 3D bbox, no 2D annotation is required).

The core challenge of mono3D is how to assign 3D targets to 2D domain with the 2D-3D correspondence and predict them afterwards.

Key ideas

Technical details

Notes