PyrOccNet: Predicting Semantic Map Representations from Images using Pyramid Occupancy Networks

June 2020

tl;dr: Predict BEV semantic map from single monocular image, or multiple streams of images.

Overall impression

From the authors of OFT, and this seems to be the natural extension of and next hot topic beyond monocular 3D object detection.

Traditional stack to generate BEV map:

Many of these tasks can benefit each other. Thus an end-to-end network to predict BEV map makes sense.

PyrOccNet ises direct supervision.

Key ideas

Technical details