Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ImagineMap: Enhanced HD Map Construction with SD Maps

Yishen Ji, Zhiqi Li, Tong Lu

TL;DR

The paper tackles mapless autonomous driving by integrating standard-definition (SD) maps as priors to jointly detect lanes, areas, and traffic elements and to infer their topology. It introduces ImagineMap, a two-stage perception-and-reasoning framework inspired by TopoMLP, with DETR-based lane and Deformable DETR-based traffic heads and topology heads that reason on upstream outputs. Extensive experiments on OpenLane-V2 show that SD-map-guided feature fusion and auxiliary BEV segmentation consistently improve lane, area, and traffic-element detection as well as topology accuracy, with ablations quantifying gains from backbone choices, BEV scaling, map-query interactions, and topology modules. These results demonstrate that SD maps can serve as practical priors to enable robust, topology-aware perception for mapless autonomous driving.

Abstract

Track Mapless demands models to process multi-view images and Standard-Definition (SD) maps, outputting lane and traffic element perceptions along with their topological relationships. We propose a novel architecture that integrates SD map priors to improve lane line and area detection performance. Inspired by TopoMLP, our model employs a two-stage structure: perception and reasoning. The downstream topology head uses the output from the upstream detection head, meaning accuracy improvements in detection significantly boost downstream performance.

ImagineMap: Enhanced HD Map Construction with SD Maps

TL;DR

The paper tackles mapless autonomous driving by integrating standard-definition (SD) maps as priors to jointly detect lanes, areas, and traffic elements and to infer their topology. It introduces ImagineMap, a two-stage perception-and-reasoning framework inspired by TopoMLP, with DETR-based lane and Deformable DETR-based traffic heads and topology heads that reason on upstream outputs. Extensive experiments on OpenLane-V2 show that SD-map-guided feature fusion and auxiliary BEV segmentation consistently improve lane, area, and traffic-element detection as well as topology accuracy, with ablations quantifying gains from backbone choices, BEV scaling, map-query interactions, and topology modules. These results demonstrate that SD maps can serve as practical priors to enable robust, topology-aware perception for mapless autonomous driving.

Abstract

Track Mapless demands models to process multi-view images and Standard-Definition (SD) maps, outputting lane and traffic element perceptions along with their topological relationships. We propose a novel architecture that integrates SD map priors to improve lane line and area detection performance. Inspired by TopoMLP, our model employs a two-stage structure: perception and reasoning. The downstream topology head uses the output from the upstream detection head, meaning accuracy improvements in detection significantly boost downstream performance.

Paper Structure

This paper contains 14 sections, 4 equations, 1 figure, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Lanesegment Detection
  4. Area Detection
  5. Traffic Element Detection
  6. Topology Lane-lane
  7. Topology Lane-Traffic
  8. Experiments
  9. Dataset and Metrics
  10. Experimental Details
  11. Lane Detection
  12. Area Detection
  13. Traffic Element Detection
  14. Topology

Figures (1)

  • Figure 1: The structure for incorporating SD map information into detection. Map encoder extracts SD map features to interact with lane queries via cross-attention. Additionally, we rasterize SD map to generate a mask, whose features are extracted by a small map backbone and concatenated with BEV features from BEV encoder for further interaction with lane queries.