# Road Genome: A Topology Reasoning Benchmark for Scene Understanding in Autonomous Driving

**Authors:** Huijie Wang, Zhenbo Liu, Yang Li, Tianyu Li, Li Chen, Chonghao Sima, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, Jun Yao, Yu Qiao, Hongyang Li  
**arXiv:** [2304.10440](https://arxiv.org/abs/2304.10440)  
**Date:** 2023-04-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.10440v1/

## Key points

- Understanding the complex traffic environment is crucial for self-driving vehicles
- Existing benchmarks in autonomous driving mainly focus on perception problems
- Limited approach to scene understanding provides insufficient information for accurate navigation
- Road Genome introduces a new benchmark that focuses on topology reasoning for scene structure
- Goal of Road Genome is to understand the relationship between perceived entities among traffic elements and lanes
- Road Genome consists of 2,000 sequences of multi-view images captured from diverse real-world scenarios
- Data is annotated with high-quality manual checks for accurate and reliable information
- Benchmark comprises three subtasks, including 3D lane detection inherited from OpenLane
- Road Genome plans to host Challenges in the near future at top-tiered venues

## Summary

Understanding the complex traffic environment is crucial for self-driving vehicles. However, existing benchmarks in autonomous driving mainly focus on perception problems, such as detecting lanelines using basic detection or segmentation methods. This limited approach to scene understanding provides insufficient information for autonomous vehicles to navigate accurately, especially without the aid of high-definition (HD) maps. For example, following the wrong traffic signal at a complicated crossroad could result in a catastrophic incident. To address this issue and go beyond perception, Road Genome (OpenLane-V2) introduces a new benchmark that focuses on topology reasoning for scene structure. The goal of Road Genome is to understand the relationship between perceived entities among traffic elements and lanes in order to better comprehend the scene structure. Built on top of prevailing datasets, this newly minted benchmark consists of 2,000 sequences of multi-view images captured from diverse real-world scenarios. Road Genome annotates data with high-quality manual checks in order to provide accurate and reliable information. The benchmark comprises three subtasks, including 3D lane detection inherited from OpenLane. Additionally, Road Genome plans to host Challenges in the near future at top-tiered venues. The expanded context reveals that Road Genome aims to tackle the limitations of existing benchmarks by focusing on topology reasoning for scene structure rather than solely relying on perception-based approaches. By investigating the relationship between perceived entities among traffic elements and lanes, Road Genome aims to provide a more comprehensive understanding of the complex traffic environment.