Learning-Deep-Learning

Deep Structured Crosswalk: End-to-End Deep Structured Models for Drawing Crosswalks

August 2020

tl;dr: Extract structured crosswalk from BEV images.

Overall impression

There are several works from Uber ATG that extracts polyline representation based on BEV maps.

This work predict deep feature maps, and use energy maximization to perform inference. Not exactly end to end.

Deep Structured Crosswalk can be directly applied to extract road boundaries. Deep Boundary Extractor is inspired by Deep Structured Crosswalk and uses conv-Snake to predict in an autoregressive fashion.

In a sense, this work is basically edge-aware semantic segmentation. The structured prediction module converts the unstructured semantic segmentation results into a structured presentation.

Key ideas

Technical details

Notes