Learning-Deep-Learning

EgoNet: Exploring Intermediate Representation for Monocular Vehicle Pose Estimation

August 2021

tl;dr: Patch-based refinement module to predict vehicle orientation based on wireframes.

Overall impression

EgoNet extracts heatmaps of object parts from local object appearances, which are mapped to the screen coordinates and then further lifted to 3D object pose.

The wireframe (or sparse 3D point cloud representating an interpolated cuboid) idea is similar to that of RTM3D and KM3D-Net and FQNet.

Key ideas

Technical details

Notes